Sage Journals: Discover world-class research

Abstract

The ventilation system is one of the essential safety systems in underground power spaces. Over the years, active ventilation has been widely employed for heat dissipation in underground power spaces. In operation, high-power equipment generates significant heat, necessitating sufficient heat dissipation for smooth and efficient functioning. The effectiveness of the ventilation system is influenced by airflow, making aerodynamics a crucial aspect of studying underground power spaces. This study establishes a comprehensive hybrid model (a combination of physical and data-driven models) representing underground power spaces. Ansys Fluent and MATLAB are used to simulate and calculate temperature fields for various structures. The physical model employs model order reduction to achieve efficient computation without compromising accuracy. For the data-driven model, a genetic neural network is developed for multifactor nonlinear optimisation to evaluate and analyse thermal behaviour within the space. The integrated hybrid model enables efficient and high-precision calculations for the underground power space’s ventilation system. The research outcomes provide a theoretical foundation for practical construction and design schemes of underground power spaces, contributing significantly to ensuring their safety and optimal functionality in real-world applications.

Keywords

Hybrid model genetic neural network underground power space temperature field

Introduction

The impact of overground power equipment on the surrounding environment, coupled with the aesthetic concerns posed by high-voltage equipment in urban settings, has been extensively studied.¹ As urban development heavily relies on energy supply, the presence of numerous overground substations can significantly influence the surrounding environment.² To address these challenges, the concept of underground power spaces has emerged, including buildings and ancillary facilities constructed below ground level within cities to accommodate power equipment,³ cables⁴ and power engineering pipelines.⁵

The ventilation system in these underground power spaces assumes a critical role in effectively managing indoor conditions. It must efficiently dissipate waste heat and eliminate toxic and harmful gases in a timely manner,⁶ while also ensuring a constant supply of fresh air for the maintenance personnel.⁷ Particularly in scenarios where high temperatures are encountered,⁸ the ventilation system needs to respond rapidly, facilitating a quick reduction in indoor temperature to enable the safe evacuation of individuals within a minute while minimising potential losses.⁹

As researchers and engineers dedicated to addressing the challenges presented by underground power spaces, the imperative trends in employing innovative approaches and methodologies. Thorough studies of the characteristics and requirements specific to these spaces guide the development of advanced ventilation systems that effectively balance the needs for heat dissipation, gas removal, and fresh air supply.

The research focuses on the design of ventilation systems with fast response capabilities to mitigate potential risks associated with high temperatures.^10,11 By prioritising the well-being and safety of individuals within these spaces, while minimising losses in critical situations, it aims to contribute to the overall resilience and efficiency of underground power spaces.

One of studies concerns in addressing a technical challenge related to the ventilation system design for underground power spaces. The primary objective is to develop a robust design methodology that effectively resolves the issue of limited feasibility in verifying the performance of underground power space systems through a large number of experiments.¹² To overcome this challenge, an innovative approach was employed, utilising the application of approximate models and simulations.¹³ To tackle this problem effectively, a substitute model was established as a viable alternative to the real-world system. The main purpose of this substitute model was to simplify the time-consuming simulation calculations involved in the optimisation iteration process. By reducing the nonlinearity of the problem, the substitute model streamlined the determination of the global optimisation point, facilitating the design process.¹⁴

The combination of physical and data-driven model, known as hybrid model, emerges as a promising approach for ventilation systems, exhibiting great potential in solving technical challenges. In several researches, a model order-reduction method is applied to develop a physical model, while the Latin Hypercube method is utilised to extract simulation data from Ansys Fluent software for fast response analysis. One of studies focuses on investigating the ventilation system of a subsurface power space using hybrid model, which enables the interactive mapping of real objects onto virtual representations in a consistent and bidirectional manner.¹⁵ Functioning as a bridge between the physical and digital fields, hybrid model technology holds broad prospects within the power industry. Its capability to map physical entities across spatial scales and throughout life cycles to the digital domain is particularly valuable. Currently, SolidWorks serves as the typical modelling tool for subsurface power spaces, while Ansys Fluent software is employed for simulating and calculating the ventilation system within these spaces.^16–18 Nevertheless, conventional finite element analysis methods have limitations due to the complex nature of the models and the need to simultaneously discretise both the space and time domains. These limitations include computational complexity and significant time consumption, normally taking hours to complete. In contrast, hybrid model technology, as a novel multiphysical field simulation platform, enables the computation of global operating parameter distributions through field computing models. Subsequently, these parameters can be utilised to drive hybrid model for device monitoring.¹⁹ The real-time nature of hybrid model technology necessitates high-speed calculations, imposing stringent requirements on computational efficiency.

In a previous study, time constraints were established to address the simulation performance of hybrid model.²⁰ The determination of the time domain error limit took into account the evaluation criteria based on the object’s requirements and application characteristics. Subsequently, the time domain error limit was integrated with model order-reduction algorithms to generate a reduced-order model that maintains high fidelity. The significance of full-scale multiphysical field model order-reduction technology in constructing hybrid model for power devices and ensuring both calculation accuracy and efficiency was emphasized.²¹ This technology plays an important role in guaranteeing accurate calculations while improving the efficiency of the model. An illustrative case based on flow-temperature field simulations, utilising the model order-reduction technology provided by Ansys Fluent,²² demonstrated substantial reductions in computation time. Compared to conventional finite element models that could take hours to compute, reduced-order models achieved computation times in mere seconds. This highlights the remarkable improvement in efficiency facilitated by model order-reduction techniques.

This paper employs the model order-reduction method to address the complexity of physical and multiphysical field coupling models. The Rom Builder Pre plugin is utilised to reduce the model order, ensuring calculation accuracy while enhancing simulation response speed and reducing hardware requirements. Through the integration of physical model and Data-driven model, the study focuses on detecting and predicting wind speed and temperature fields within a subsurface power space ventilation system, considering existing equipment conditions. Subsequently, an analysis of the observed patterns and behaviours is conducted, leading to the formulation of a tailored program for the ventilation system. This research presents a theoretical foundation for practical construction designs of subsurface power spaces. To solve the aforementioned technical issue, the modelling of the underground power space is established using SolidWorks. Subsequently, the ventilation system of the underground power space is simulated and calculated using the finite element software Ansys Fluent.²³ This analysis encompasses the calculation of air speed and temperature fields, followed by the summarisation of observed patterns. Furthermore, this paper presents a comprehensive design scheme for the ventilation system based on the simulation and calculation results using an integrated hybrid model. The outcomes of this study provide a valuable theoretical basis for the practical implementation of construction design schemes for underground power spaces.

Hybrid model

Hybrid model approach, encompassing a combination of physical and data-driven model, serves as a prevalent approach for simulating and analysing subsurface power space ventilation systems. The physical model utilises a three-dimensional (3D) model that integrates historical characteristic parameters.²⁴ This integration enables real-time visualisation and responsiveness to the temperature field of the ventilation system.²⁵ On the other hand, the data-driven model processes and utilises real-time parameters generated during system operation to predict the temperature field.

In this study, the subsurface power space ventilation system is selected as the focal point for establishing a hybrid model within the Internet of Things (IoT) platform. The overall architecture of the hybrid model is depicted in Figure 1, illustrating the interconnected components and their relationships.

(1) The physical layer of the system encompasses various essential attributes. It encompasses service data related to heat-generating equipment, such as their operational characteristics and specifications. Additionally, it includes the working parameters of the ventilation equipment, which are crucial for ensuring proper airflow and temperature regulation within the subsurface power space.²⁶

(2) The cloud infrastructure, generated by the infrastructure as a service layer, plays a critical role in managing the data originating from the foundation support layer and hybrid model. Through the utilisation of data management tools within the platform as a service layer, the cloud infrastructure facilitates the seamless transmission, high-speed processing, storage, and analysis of massive datasets derived from electrical equipment. This infrastructure empowers efficient data handling and enables comprehensive insights into the performance and behaviour of the electrical equipment.²⁷

(3) Utilising a combination of real-scene modelling, visualisation techniques, multiphysical field simulation, and big data analysis, the process of hybrid model modelling and simulation is undertaken. By creating hybrid model based on the device specifications found in the physical layer, a virtual representation of the physical devices is achieved, visually capturing their characteristics and attributes. The application of Fluent software enables multiphysical field simulation, specifically focusing on the ventilation system within the subsurface power space. Through this simulation, real-time observations of temperature field variations become possible. Furthermore, leveraging feature extraction methodologies and iterative training of the BP neural network, scientific predictions of the temperature field can be made, allowing for assessments of system reliability and performance. This integration of simulation, analysis, and prediction techniques serves to enhance the understanding and decision-making capabilities related to the subsurface power space ventilation system.^28,29

(4) The utilisation of a hybrid model software platform, based on Software as a Service (SaaS) architecture, offers a comprehensive solution for implementing hybrid model technology throughout the entire equipment life cycle. This platform enables the optimisation of system design and provides the capability to evaluate and predict the operational state of the system. By leveraging the features and functionalities of the hybrid model software platform, users can efficiently navigate through various stages of equipment management, including design, deployment, operation, and maintenance. The platform empowers decision-makers to make informed choices, enhance system performance, and ensure efficient and reliable operation throughout the equipment’s life cycle.³⁰

Figure 1.

Overall architecture of the hybrid model based on the IoT.

The above space introduces the general situation of the parameter optimization of the underground power space ventilation system based on the Internet of Things. Figure 2 shows the specific process of the hybrid model used in this study.

Figure 2.

Flow chart of the hybrid model.

Physical model

Within this subsection, the construction of the subsurface power space model takes place, followed by the utilisation of Fluent software for simulating and calculating the various inputs and outputs of the ventilation system. The focus lies on the comprehensive analysis of the wind speed field and temperature field. Achieving a close match between the simulated data and the actual data necessitates meticulous meshing of the model to enhance calculation accuracy. However, dense meshing and the inclusion of multiple parameters lead to prolonged simulation times. To address this challenge, the model order-reduction method employs Latin hypercube sampling to extract the necessary data, which is subsequently fed into the Rom Builder Pre plugin for computation. This approach maintains a level of accuracy similar to that achieved through Fluent simulation while ensuring broad applicability. By capturing the system’s energy dynamics through a low-dimensional approximation of multidimensional physical processes over time, this method effectively reduces the computational burden in terms of Central Processing Unit (CPU) load, dimensions, quantity, and time. During the system’s actual operation, the reduced-order models exhibit rapid analysis and response capabilities to the collected data, thus significantly enhancing the system’s reliability and performance. This approach enables efficient decision-making and supports real-time monitoring and control, contributing to the overall robustness of the subsurface power space ventilation system.

Singular value decomposition

The singular value decomposition (SVD) algorithm, employed in the realm of machine learning, serves as a matrix data compression technique. It effectively approximates the original matrix through a factor decomposition formula, resulting in the derivation of an order-reduced system $\hat{H} (s)$ from the initial system. $H (s)$ By decomposing the matrix into singular values, the SVD algorithm enables the representation of complex data in a more compact and simplified form, while preserving essential characteristics and minimising information loss.

H : {\begin{matrix} \overset{\cdot}{x} (t) = Ax (t) + Bu (t) \\ y (t) = Cx (t) \end{matrix}

(1)

\hat{H} : {\begin{matrix} \overset{\cdot}{\hat{x}} = \hat{A} \hat{x} (t) + \hat{B} \hat{u} (t) \\ \hat{y} (t) = \hat{C} \hat{x} (t) \end{matrix}

(2)

Here, $A \in R^{n \times n}$ , $B \in R^{n \times n_{u}}$ , $C \in R^{n_{y} \times n}$ , $\hat{A} = W^{T} AW$ , $\hat{B} = W^{T} B$ and $\hat{C} = CV$ ( $V, W \in R^{n \times r}$ and $W^{T} V = I_{r}$ ). $H$ is denoted by $\hat{x} \in R^{n}$ , and the target order-reduction system $\hat{H}$ is denoted by $\hat{x} \in R^{r}$ . $W$ and $V$ represent projected left and right vectors in order-reduction operations, respectively.

min_{\hat{H} (s), (stable, r)} J = ‖ H (s) - \hat{H} (s) ‖_{H_{2}}

(3)

where the Grampian matrices P and Q, which are both controllable and observable, fulfil the requirements of the Lyapunov equation:

{\begin{matrix} AP + P A^{T} + B B^{T} = 0 \\ A^{T} Q + QA + C^{T} C = 0 \end{matrix}

(4)

The base matrix $S$ of the left eigenspace and the base matrix $T$ of the right eigenspace of the matrix $G_{PQ} = PQ$ was calculated. The SVD method can be used to obtain an order-reduction system with $r$ orders using $S$ and $T$ through the following steps. First, it assumes that $λ_{1}^{2}$ , $λ_{2}^{2}$ , … , $λ_{n}^{2}$ are $n$ nonrepeating eigenvalues of the matrix $G_{PQ}$ arranged in descending order and that $S$ and $T$ are matrices composed of the left and right eigenvectors corresponding to the eigenvalues $λ_{1}^{2}$ , $λ_{2}^{2}$ , … , $λ_{n}^{2}$ of the matrix $G_{PQ}$ :

{\begin{matrix} S^{T} G_{PQ} = diag (λ_{1}^{2}, λ_{1}^{2}, \dots, λ_{n}^{2}) S^{T} \\ G_{PQ} T = Tdiag (λ_{1}^{2}, λ_{1}^{2}, \dots, λ_{n}^{2}) \end{matrix}

(5)

Let $S = [S_{B} \begin{matrix} S_{S} \end{matrix}]$ and $T = [T_{B} \begin{matrix} T_{S} \end{matrix}]$ , where $S_{B}$ and $T_{B}$ denote the matrices composed of the left and right eigenvectors corresponding to the first $r$ eigenvalues of the matrix $G_{PQ}$ . The subscript $S$ indicates the matrix consisting of the eigenvectors corresponding to the remaining eigenvalues.

Let $E = S_{B}^{T} T_{B}$ . $E$ can be decomposed by SVD:

E = U_{E} \sum R_{E}

(6)

Then, we have $W = S_{B} U_{E} \sum^{\frac{1}{2}}$ , $V = T_{B} R_{E}^{T} \sum^{\frac{1}{2}}$ .

By substituting equation (3) into equation (2), it becomes possible to derive the system matrix of the reduced-order model. Through this substitution process, the complexities of the original equations are effectively condensed, providing a more manageable and insightful framework for studying the system dynamics.

Date-driven model

The research and analysis conducted in this study focused on a specific project, where variations in a single influencing factor were examined while keeping other factors constant. However, it is important to acknowledge that the temperature field distribution in a power space is influenced by multiple factors simultaneously. Therefore, it becomes vital to consider a comprehensive set of factors when determining the optimal results.³¹ Due to practical constraints, conducting precise and large-scale experimental verification in an actual underground power space system is infeasible.³² To address this limitation, an approximate model simulation method was employed, enabling the derivation of an optimisation scheme through analysis of the simulation results.³³ The approximate model method encompasses experimental design, data fitting and predictive modelling techniques to establish alternative models that represent real-world scenarios.³⁴ This approach simplifies the time-consuming simulation calculations through an optimisation iteration process, reducing nonlinearity in the problem and aiding in the identification of the global optimisation point.³⁵ In order to achieve comprehensive optimisation, a BP artificial neural network model was developed, followed by the utilisation of genetic algorithms to determine the optimal results. This integrated approach offers a more robust and effective optimisation process, taking into account the complex interplay of various factors in the underground power space system.

Modelling procedures

To create the wall model in SolidWorks, a cuboid shape was constructed with inner dimensions of 16 × 16 × 12 m and a wall thickness of 1 m. Various design schemes were proposed by manipulating different parameters, including the location of the air inlet, the number and placement of air outlets, the air speed at the outlets, the surface area of equipment, and the power of the equipment. To establish the air inlet configuration, a distance of 0.3 m was maintained between the lower part of the inlet and the ground plane of the wall. In addition, two air inlets with dimensions of 1.414 × 1.414 m were positioned at different distances of 4, 8 and 12 m from the central line of the plane. These specific design variations allowed for the exploration and analysis of different model distributions to assess their impact on the ventilation system within the constructed wall model.

Air out experimental methodology

The wall model in the simulation consisted of vents, each with dimensions of 1 × 1 m, placed on the upper surface of the wall. To explore different scenarios, three schemes of air outlets were applied: The first scheme involved two air outlets located at a distance of 0.5 m from the outer wall. In the second scheme, three air outlets are positioned at interval of 1 m, while in the third scheme, four air outlets were placed at interval of 1.5 m. For the location of the air outlets, variations are considered with two outlets separated by distances of 4, 8 and 12 m; three outlets spaced at intervals of 2, 4 and 6 m; and four outlets set at intervals of 1, 2 and 3 m. These distributions are denoted by the labels 1, 2 and 3, representing ‘a small distance’, ‘a moderate distance’ and ‘a large distance’ for different numbers of air outlets, respectively (see Table 1). The air speed at the exhaust outlet is set at 5, 6 and 7 m/s, respectively. To investigate the impact of heating devices, a rectangular object was used to represent the heating equipment, with varying surface areas. The sizes and surface areas of the devices were set as 8 × 4 × 6 (176 m²), 8 × 6 × 6 (216 m²) and 8 × 8 × 6 (272 m²). The heating power of the equipment is set at 100, 200 and 300 kW. The independent variables and the design levels are summarised in Table 1. Figure 3 shows a simplified model of the underground power space, where several independent variables have specific values: x1 = 4 m, x2 = c, x3 = 1.5 m, x4 = 4 and x6 = 176 m².

Table 1.

Optimisation design parameters of the ventilation system.

Independent variable	Design level
Change of air inlet location –x1	4, 8, 12
Change of air outlet location –x2	Small, moderate, large
Distance of the air outlet location to the wall location –x3	0.5, 1, 1.5
Number of air outlet –x4	2, 3, 4
Air speed of air outlet –x5	5, 6, 7
Surface area of equipment –x6	176, 216, 272
Heating power of equipment –x7	100, 200, 300

Figure 3.

Simplified model of an underground power space.

Physical model and boundary conditions

The physical model in the analysis consists of a specified geometry representing the object or system under study or research. It includes the relevant dimensions, shape and structural properties of the components involved. In addition to the physical model, the boundary conditions are essential parameters that define the interactions between the model and its surroundings. The boundary conditions establish constraints, such as fixed supports, applied forces or prescribed displacements, which influence the behaviour of the physical model within the analysis. By accurately applying the physical model and boundary conditions, the analysis can simulate and predict the response of the system to different operating conditions or external forces.

Mesh generation

Figure 4 illustrates model 083, serving as a representative example to elucidate the calculation procedure. A mesh is generated encompassing the entire model, ensuring comprehensive coverage. To uphold the accuracy of the calculations, a portion of the mesh is specifically encrypted, thereby enhancing precision. The mesh comprises approximately 1.1 million elements, demonstrating the level of detail employed in the computational analysis.

Figure 4.

Mesh distribution for the ventilated building.

Governing equation, numerical model and boundary conditions

The calculations for the ventilation system are conducted by applying the fundamental governing equations that describe fluid behaviour. These equations include the continuity equation, which ensures the conservation of mass, as well as the momentum and energy equations.³⁶ Since the air in the system operates at high velocities, it is considered to be in a turbulent state.³⁷ Therefore, the transport equation for turbulent kinetic energy is also taken into account. To model the turbulence, the standard k-ε model is used, while treating the gas as a Boussinesq flow.³⁸

In the model, a combination of natural air inlet and mechanical air exhaust is employed as the underground ventilation mode. The air is expelled from the system through the exhaust outlet with a velocity of −7 m/s. The power generation equipment generates heat with a dissipation rate of 300 kW. Based on this configuration, the cooling equipment is assigned a heat flow boundary of 1704.55 W/m², while the walls are designed as heat flow walls. The building wall, constructed with concrete, is thermally isolated, whereas other components are fabricated using Q235A steel.

Solution settings and convergence criteria

(1) Computing Analysis

In the computing analysis, the simulation is carried out under a standard atmospheric pressure of one atmosphere, which is equivalent to 101,325 Pascals (Pa). The gravitational acceleration in the Y direction is defined as −9.81 m²/s. This value represents the acceleration due to gravity, ensuring that the simulation accurately reflects the physical reality.

(2) Setting of Solution Control Options

To effectively handle the coupling of pressure and velocity, the simulation employs the SIMPLEC algorithm. For the computation of pressure difference, the simulation uses the PRESTO! scheme. By employing the PRESTO! scheme, the simulation can effectively handle pressure variations and accurately model the behaviour of the fluid system. To accurately simulate the momentum, turbulent kinetic, and turbulent dissipation rate equations, the simulation utilises the first-order upwind scheme.

(3) Judgment of Convergence Criteria

In the simulation, convergence criteria are established to determine when the solution has reached a stable and consistent state. The continuity equation, the velocity equations in the X/Y/Z directions, the energy associated with turbulent motion, and the dissipation rate equation are considered converged when the discrepancy between successive iterations reaches a value of 0.001. Furthermore, the energy equation has a more stringent convergence criterion of 1 × 10⁻⁶. This higher precision criterion ensures that the simulation accurately models the energy transfer processes and provides reliable temperature predictions.

Multivariate optimisation sampling of the subsurface power space ventilation system

Recently, Latin hypercube sampling (LHS) has developed as a widely employed method for experimental design. Rooted in probability theory, mathematical statistics, and linear algebra, LHS facilitates the systematic arrangement of sample data points, thereby mitigating the impact of random errors. By reducing the number of required experimental samples, LHS not only saves computation time but also ensures the accuracy of fitting and experimental design. In the context of this study, the optimisation of the ventilation system involves multiple factors, a total of 100 test plans were initially generated using LHS. These test plans becomes the foundation for subsequent research and analysis, particularly in conjunction with the BP neural network. By leveraging constructed test plans, the study aimed to obtain valuable insights and draw meaningful conclusions regarding the performance of the ventilation system. The sampling principle was as follows: Under the premise that there is only one test plan in a hyperplane, 100 sets of test plans are randomly selected.

The LHS algorithm follows specific steps to ensure an effective sampling process in this study:

Step 1: Equally divide the seven parameters into three factors.

Step 2: Randomly assign a numerical value to each parameter, representing a different level.

Step 3: Combine the seven parameters in a way that ensures each hyperplane has only one parameter combination.

Step 4: Generate the initial population based on the previous steps.

Step 5: Use the genetic algorithm to optimize the results and output the sampling outcomes.

When selecting the interval for process parameters, certain requirements must be met:

(1) The interval should encompass a wide range of scenarios to capture various situations.

(2) The interval range should not be excessively large to avoid compromising accuracy due to a significant distance between sample points.

(3) Consider the actual case and keep a balance by choosing a reasonable interval that combines the aforementioned points.

The design level of the optimised ventilation system using LHS is presented in Table 1, and the samples drawn with the LHS method are listed in columns 2–7 of Table 2.

Table 2.

Sampling and calculation results obtained using the Latin hypercube sampling (LHS) method.

Number	Location of air inlet	Location of air outlet	Distance of air outlet to the wall location (m)	Number of air outlet	Air speed of air outlet	Surface area of equipment	Heating power of equipment (kW)	Surface temperature of equipment (°C)	Surface temperature of equipment at a distance of 1 m (°C)	Temperature of air outlet (°C)
1	12	3	1	4	5	272	200	144.01	39.49	38.89
2	8	3	1	2	7	176	300	149.63	47.88	48.12
3	4	1	0.5	3	5	216	100	97.11	34.47	36.2
4	4	2	1	2	7	272	100	95.31	34.94	36.62
5	12	3	1.5	4	6	176	100	108.3	36.15	34.06
6	4	2	0.5	2	5	176	200	221.15	38.39	46.76
7	12	2	1.5	3	6	176	300	234.06	39.45	44.1
8	8	1	1	2	5	176	100	134.69	34.7	38.79
9	12	3	1.5	3	5	272	100	101.39	36	36
10	8	3	1.5	4	6	272	200	121.01	38.18	37.52
11	8	2	1	4	7	176	300	204.9	41.26	39.55
12	4	1	1	2	6	272	300	219.33	41.96	50.73
13	4	3	1	2	7	216	100	100.7	34.66	36.57
14	12	1	1	2	5	272	200	122.37	44.45	46.95
15	4	1	1	3	7	176	100	99.22	33.89	34.57
16	8	3	1	4	6	272	200	120.72	37.99	37.57
17	12	3	1	4	7	176	100	103.75	35.7	33.64
18	12	1	0.5	2	7	176	300	259.54	39.07	47.9
19	4	1	1.5	3	5	216	200	158.87	36.99	41.64
20	4	3	1	4	6	216	200	126.38	35.98	37.52
21	4	2	0.5	3	6	272	200	148.94	36.59	39.82
22	8	2	0.5	4	7	176	100	100.06	34.95	33.61
23	4	1	1.5	3	6	272	200	107.74	42.73	44.15
24	8	1	1	4	5	216	300	189.92	39.38	42.9
25	8	2	1	3	6	216	100	98.24	35.4	35.35
26	12	2	0.5	3	5	176	200	186.71	36.83	41.38
27	4	2	0.5	3	5	176	100	112.41	35.05	36.16
28	12	2	1	2	5	272	100	108.87	35.05	38.77
29	4	2	1.5	3	6	272	200	151.77	36.39	39.88
30	12	2	1	4	5	272	300	190.02	41.09	42.9
31	8	1	0.5	3	7	216	300	187.3	39.44	42.43
32	8	2	1	3	6	216	200	150.11	37.46	39.91
33	12	1	1.5	4	6	216	300	112.67	31.08	31.96
34	4	2	0.5	4	7	272	300	163.3	37.57	39.51
35	8	1	1.5	4	7	272	200	108.05	36.75	36.5
36	12	3	0.5	3	7	216	200	155.7	39.25	38.76
37	8	3	0.5	2	7	272	100	95.06	35.28	36.58
38	4	1	1	2	5	272	200	124.52	42.84	46.96
39	8	3	0.5	4	5	272	100	84.98	35.52	34.81
40	8	1	0.5	3	7	216	200	143.17	37.44	38.5
41	4	3	0.5	4	6	216	100	81.9	33.84	34.11
42	4	2	1	2	5	216	100	177.34	57.57	60.59
43	12	3	1.5	2	6	216	300	171.77	53.52	51.06
44	12	1	0.5	4	5	272	300	188.72	40.97	42.71
45	12	1	1.5	2	7	176	300	258.81	38.91	47.97
46	8	2	1	3	5	216	100	103.93	34.97	36.31
47	4	2	1	3	6	176	200	106.86	33.85	35.17
48	4	3	0.5	2	5	216	300	159.29	52.3	54.98
49	12	1	1.5	3	6	216	200	159.56	37.99	39.8
50	12	1	1	2	7	216	100	110.77	35.39	36.58
51	12	2	1.5	2	6	176	100	127.32	34.64	37.43
52	8	1	1.5	4	7	272	300	148.68	40.38	39.45
53	4	3	0.5	2	7	176	200	187.41	37.46	42.33
54	8	1	1.5	3	7	216	300	185.59	40.03	42.56
55	8	1	0.5	2	5	176	300	108.74	55.61	54.91
56	4	1	1	2	6	176	100	120.86	34.97	37.6
57	8	3	1.5	2	7	176	100	104.4	35.59	36.61
58	12	2	0.5	2	7	176	300	152.29	48.89	48.06
59	8	3	1.5	4	7	216	100	85.79	34.45	33.65
60	4	1	0.5	2	7	176	100	114.39	34.9	36.59
61	4	3	0.5	3	6	272	100	94.02	34.18	35.3
62	4	1	0.5	4	5	272	300	192.94	38.26	43.01
63	8	2	0.5	4	6	176	100	106.53	35.18	34.06
64	12	1	1	3	5	216	300	256.2	40.7	50.43
65	8	2	1.5	3	7	216	100	87.09	35.26	34.59
66	12	2	1.5	4	6	176	300	228.04	42.81	40.89
67	12	1	0.5	2	6	272	100	105.66	35.28	37.44
68	12	1	1.5	4	5	176	100	116.78	37.31	34.88
69	4	3	1.5	4	6	216	200	126.5	35.95	37.51
70	12	1	1.5	2	7	272	200	155.04	37.6	42.19
71	8	2	1.5	4	5	176	300	221.56	38.99	43.2
72	8	3	0.5	3	5	272	200	141.17	37.54	41.38
73	8	1	0.5	4	7	272	300	148.62	40.59	39.37
74	12	3	1	3	5	272	300	197.71	39.57	46.71
75	8	3	0.5	4	7	216	300	174.04	39.8	39.47
76	8	2	0.5	3	6	272	300	170.46	39.69	44.25
77	12	2	1	2	5	216	200	158.95	44.21	47.1
78	4	3	1	4	7	216	100	58.31	33.64	33.69
79	4	2	1	3	5	176	200	180.07	37.38	41.64
80	4	2	1	2	5	216	100	116.57	35.13	38.79
81	8	3	0.5	4	7	272	200	113.79	38.29	36.53
82	12	2	1.5	4	6	176	300	215.59	44.66	41.06
83	4	3	1.5	4	7	176	300	192.16	38.07	39.2
84	4	2	0.5	3	6	216	100	88.45	34.46	35.29
85	4	3	1.5	3	6	176	200	165.31	37.6	39.84
86	8	2	1	3	5	216	300	160.69	36.93	41.49
87	12	3	0.5	4	6	272	300	182.74	43.05	41.12
88	12	1	1	2	5	216	200	107.19	47.69	46.84
89	8	3	1.5	2	6	216	300	119.61	49.73	51.01
90	4	3	1.5	3	5	216	200	159.11	37.08	41.56
91	12	3	1.5	2	6	176	200	169.78	37.34	43.97
92	8	2	1	4	7	272	300	147.29	40.38	42.42
93	4	3	1.5	2	7	272	200	110.2	37.23	42.26
94	12	2	0.5	2	7	176	200	189.2	39.84	44.19
95	8	1	1.5	3	6	176	300	229.42	35.74	45.45
96	4	3	1	4	6	216	100	79.55	33.76	34.11
97	12	3	1.5	3	6	176	300	210.2	42.62	44.21
98	8	1	0.5	3	5	272	200	140.23	37.51	43.43
99	12	1	1.5	4	5	272	200	144.98	39.78	39.89
100	12	3	1.5	2	7	272	100	105.82	33.02	36.65

Results and discussions

The results achieved clearly showcase the prowess of the hybrid model. Unlike conventional simulation approaches, the hybrid model seamlessly combines simulation calculations, data processing, and model predictions, leading to a marked enhancement in system performance. Research findings based on the hybrid model reveal that by optimizing factors such as ventilation system vent positions, quantities, and vent speed parameters in the context of the underground power space’s structure and equipment cooling evaluation, superior temperature control can be attained for the heating equipment within this underground power space. Using the temperature field as a reference, this optimisation increases airflow distribution within a specific range, thereby bolstering the overall reliability of the system.

Analysis of simulation results

By using sophisticated finite element software, a comprehensive simulation of the underground power space system is created, encompassing essential parameters like space structure, initial equipment surface and environmental temperatures, and ventilation system operating speed, with the goal of closely emulating real-world operating conditions. This simulation method enables us to analyse the most significant fluid velocity and temperature trends within the ventilation system. It also can investigate both the impact of natural vents and mechanical vents represented by the 083 model on airflow velocity within the underground space, and also examine the effects of this ventilation system on the temperature distribution range and diffusion characteristics of the heating equipment within the underground power space.

Analysis of the velocity field

Figure 5 illustrates the comprehensive airflow patterns and velocity distribution across the building once the ventilation system attains a steady state. This diagram visually represents both the direction and intensity of the air movement throughout the building. The size relativity of the speed values is analysable based on distinct colours. Observing the numerical range of the velocity field, the maximum speed achieved by the ventilation system is up to 9.32 m/s, slightly surpassing the initial vent speed of 7 m/s. This occurrence arises due to the air flow movement’s sensitivity to temperature gradients induced by the heating equipment, resulting in the elevated maximum speed beyond the preset initial value, aligning with established engineering practices and further validating the simulation results’ reliability. Considering the velocity field’s distribution range, the average velocity near by the exhaust vent is slightly lower than that at the inlet. Concurrently, the velocity distribution near the exhaust outlet exhibits scattering in comparison to the inlet. It’s noteworthy that, within a specific space, the ventilation system’s efficacy doesn’t proportionally increase with the quantity of exhaust vents. Hence, judicious selection of exhaust vents can enhance economic benefits while ensuring optimal ventilation effectiveness.

Figure 5.

Vectograph of the overall air velocity.

Analysis of the temperature field

Figures 6 and 7 present temperature distributions in the X-m section and equipment surface, respectively, providing visual representations of temperature dispersion in these regions. By scrutinising definedpatterns, valuable insights into temperature variations across different positions within the room, spanning from 38°C to 302°C, can be gleaned. The temperature field distribution in the X-m plane (as shown in Figure 6) reveals a diffusion trend toward the exhaust vent in alignment with established engineering norms, further solidifying the simulation results’ reliability. Analysing the temperature fields shown in Figures 6 and 7, it reveals that lower temperatures occur near the vents, with notable temperature disparities between the upper and lower sections of the heating equipment. The overall temperature distribution indicates that proximity to the vent corresponds to lower temperatures, particularly in the areas directly facing the vents, where temperatures even approach room temperature. The temperature field changes observed in Figures 6 and 7 mirror the velocity field changes detailed in Figure 5 in the preceding section – specifically, the dispersal of temperature distribution near the exhaust outlet relative to the inlet. In a specific spatial context, it’s important to note that the ventilation system’s effectiveness doesn’t linearly increase with the number of exhaust vents, necessitating comprehensive research in the overall system.

Figure 6.

Pattern of temperature field at the X-m section.

Figure 7.

Pattern of temperature field on the equipment surface.

Results of physical model

The physical model approach offers a comprehensive mapping of multiple physical fields corresponding to the real operating conditions of the subsurface power space. First of all, a 3D spatial geometry model is constructed, encompassing key components such as walls, heat-generating equipment, and basins. Fluent software is then used to create a date-driven model incorporating the system’s prominent parameters, which is later merged with the geometric model. The simulation results are visualised in a 3D representation, establishing a physical model for the subsurface power space ventilation system.

To achieve the coupling of the multiphysical flow and thermal fields in the ventilation system, the Latin Hypercube method is employed to generate 100 sets of samples. These samples are added and calculated to facilitate order reduction. The temperature field in the fluid region can be therefore obtained automatically. By comparing the temperature fields driven by the feature parameters of model 083, derived from both Fluent simulations and the Order-reduced model, the differences of the result in these two methods are compared. If the relative error remains below 3%, it can be concluded that the impact on the study of the subsurface power space ventilation system is negligible, thus ensuring the accuracy of the reduced-order model’s calculations. Integrating the model order-reduction model into physical model technology offers the flexibility to train and explore multiple parameter samples, enabling variations in feature parameters across a wider range. This approach facilitates the rapid generation of simulation results, allowing for real-time analysis of fields and observation of temperature changes in the basin.

Comparison of fluent and order-reduced models

The comparison between the predicted and actual data, as depicted in Table 3 and Figure 8, involves temperature field data obtained from simulations based on 12 sets of randomly selected input parameters. This comparison serves to reveal the margin of error between the model trained using the physical model and the observed data. The analysis of these findings indicates that the deviation in predicted values displays irregular fluctuations; however, the maximum deviation in predictions remains below 0.001%, demonstrating a remarkable agreement with the results obtained from Fluent simulations. Consequently, the errors in prediction and simulation have minimal effect within this study, and this slight discrepancy exerts no discernible impact on the subsequent analysis, thereby ensuring the high reliability of the predictions.

Table 3.

Reduced-order models input parameters.

Number	1	2	3	4	5	6	7	8	9	10	11	12
S1 (°C)	1160	1150	1145	1100	1050	1030	1080	1000	1135	1140	950	980
V1 (m/s)	0.500	0.480	0.470	0.400	0.380	0.360	0.390	0.350	0.450	0.370	0.320	0.340

Figure 8.

Comparison of errors for the Fluent and reduced-order models.

Analysis of the prediction model

The real-time interface showcasing the simulation results for the physical model is illustrated in Figure 9. This interface reveals the creation of a physical model, as depicted on the left side of the Figure. The numerical display range for the simulation predictions is determined based on the actual operating parameters of the power space ventilation system. By using the display interface, input parameters can be adjusted through a slider, leveraging the established data model. This adjustment impacts the temperature field’s value range and diffusion characteristics, as demonstrated on the right side of Figure 9. This interactive capability significantly enhances the flexibility and usability of the physical model. It supports the exploration of diverse scenarios, the construction of the physical model, the fine-tuning of input parameters, and the output of corresponding cloud maps, ultimately facilitating a comprehensive understanding of temperature dynamics.

Figure 9.

Prediction model.

Results of the data-driven model

Using the inputs and outputs of the subsurface power space ventilation system, a mathematical model is developed to understand its physical characteristics and perform order-reduction calculations. By training the model with a portion of the data, an optimal mathematical model is obtained. The remaining data serves as input to the optimized model and enables mapping of the system’s operation and generating new output information.

To assess the performance of the mathematical model, 100 sets of sample data are applied for simulation calculations using Fluent software. The results of simulation calculation, including equipment surface temperature, equipment surface temperature at a distance of 1 m, and air outlet temperature, are presented in columns 9–11 in Table 2. Subsequently, these results carry out the optimisation using the development of neural network and genetic algorithms.

Figures 10 and 11 illustrate the analysis of the absolute error and relative error, respectively, in the calculation performed by the BP neural network model. The absolute error remains within 5.0, and the relative error falls within 15%, demonstrating the reliability and effectiveness of the BP neural network model. This finding provides evidence of a strong correlation between the various input parameters and the system’s temperature.

Figure 10.

Absolute error in the optimisation results of the BP neural network model.

Figure 11.

Relative error in the results obtained using the BP neural network model.

Conclusions

This paper examines the planning and design of underground power space ventilation systems, establishing the safety design and safety evaluation of these systems. Using a comprehensive hybrid model that integrates Ansys Fluent and MATLAB, the study simulates and calculates the air speed field and temperature field within underground power spaces. It also summarises universal principles and completes the calculation and simulation results, providing a theoretical foundation for practical construction design schemes.

The study’s main conclusions are as follows:

Temperature and Flow Field Distribution: Simulations with various wind speeds and conditions reveal distinct distributions of temperature and flow fields. The results show that selecting an optimal number and location for air inlets effectively reduces temperatures. In particular, positioning the inlets higher in the space promotes better ventilation.

Model Order Reduction and Simulation: To simplify response analysis, a model order reduction method was adopted for the physical model. This approach allows for efficient computation without sacrificing accuracy. The Latin hypercube method was used to extract simulated data from Ansys Fluent, enabling rapid analysis of the complex dynamics of the ventilation system.

Multi-Factor Optimisation: By employing MATLAB software, a multi-factor optimisation model based on BP neural network was developed. This model effectively addresses the complex challenge of multi-factor nonlinear optimisation in underground power space systems. The findings offer significant insights for applying BP neural networks to other ventilation systems.

Prediction and Analysis: Utilising the established hybrid model, the study comprehensively predicts and assesses the impact of various factors on temperature fields within underground power spaces. This approach provides valuable reference information for the operation and design of ventilation systems.

The study’s findings have practical implications for the construction and design of underground power space ventilation systems. The results contribute to improved safety and optimisation in real-world applications.

Footnotes

Acknowledgements

We also wish to acknowledge the many participating industry partners, without whom the project would not be possible.The PowerChina Hebei Electric Power Engineering Co, Ltd. and the School of Mechanical Engineering, Shijiazhuang Tiedao University, participated in the model construction and simulation, as well as the optimization and prediction of ventilation system parametersand was a major contributor in writing the manuscript. All authors read and approved the final manuscript.

Handling Editor: Assunta Andreozzi

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the PowerChina Hebei Electric Power Engineering Co, Ltd. and the School of Mechanical Engineering, Shijiazhuang Tiedao University.

Research ethics

The research has been conducted with the highest standards of rigour and integrity.

The article/chapter/book/case study is original.

The work has not been submitted elsewhere and is not under consideration with any other publication. See our preprint and conference paper policies for exceptions.

The work does not include libellous, defamatory or unlawful statements.

Permission has been cleared for any third-party material included.

Proof of consent has been obtained for any named individuals or organisations.

Authorship has been agreed prior to submission and no one has been ‘gifted’ authorship or denied credit as an author (ghost authorship).

ORCID iD

Kang Su

References

Sadeghi

Ghandehariun

Thermodynamic analysis and optimization of an integrated solar thermochemical hydrogen production system. Int J Hydrogen Energy 2020; 45(53): 28426–28436.

Kazemi

Nor

MIM

Teoh

WH.

Thermodynamic and economic investigation of an ionic liquid as a new proposed geothermal fluid in different organic Rankine cycles for energy production. Energy 2020; 193: 116722.

Beygzadeh

Khalilarya

Mirzaee

Thermodynamic comparison of two novel combined systems based on solar loop heat pipe evaporator. Energy 2020; 206: 118145.

Paweł

Monika

Venkata

, et al. Multiobjective optimization of underground power cable systems. Energy 2021; 215: 119089.

Klimenta

Tasić

Perović

, et al. Eliminating the effect of hot spots on underground power cables using cool pavements. Electr Eng 2019; 101: 1295–1309.

Moradi

Saffarian

Behbahani-Nejad

Experimental study of an air humidity absorption cycle based on the MHI. J Therm Anal Calorim 2020; 139(6): 3613–3621.

Wang

, et al. High-efficiency solar evaporator prepared by one-step carbon nanotubes loading on cotton fabric toward water purification. Sci Total Environ 2020; 698: 134136(C).

Neumayr

Knabben

Varescon

, et al. Comparative evaluation of a full-and partial-power processing active power buffer for ultra-compact single-phase DC/AC converter systems. IEEE J Emerg Sel Top Power Electron 2020; 9(2): 1994–2013.

McGill

Altchenko

Hamilton

, et al. Complex interactions between climate change, sanitation, and groundwater quality: a case study from Ramotswa, Botswana. Hydrogeol J 2019; 27(3): 997–1015.

10.

Belhocine

Shinde

Patil

Thermo-mechanical coupled analysis based design of ventilated brake disc using genetic algorithm and particle swarm optimization. JMST Adv 2021; 3(3): 1–14.

11.

Belhocine

Omar

WZW

. Computational fluid dynamics (CFD) analysis and numerical aerodynamic investigations of automotive disc brake rotor. Int J Interact Des Manuf 2018; 12(2): 421–437.

12.

Soo

Vuong

Geun

, et al. Design and electromagnetic analysis of an induction-type coilgun system with a pulse power module. IEEE Trans Plasma Sci 2019; 47(1): 971–976.

13.

Heydarian-Forushani

Golshan

HEM

Shafie-khah

, et al. A comprehensive linear model for demand response optimization problem. Energy 2020; 209: 118474.

14.

Zhang

Tang

Chen

Linear substitute model-based uncertainty analysis of complicated non-linear energy system performance (case study of an adaptive cycle engine). Appl Energy 2019; 249: 87–108.

15.

Xiang

Pan

Luo

A new model updating strategy with physics-based and data-driven models. Struct Multidiscip Optim 2021; 64(1): 1–14.

16.

Guangwei

Shaonan

, et al. Research on wind direction measurement of wind turbine based on fluid simulation. J Phys Conf Ser 2023; 2441(1): 012059.

17.

Belhocine

Abdullah

IO.

Finite element analysis (FEA) of frictional contact phenomenon on vehicle braking system. Mech Based Des Struct Mach 2020; 50(9): 1–36.

18.

Afzal

Soudagar

Belhocine

, et al. Thermal performance of compression ignition engine using high content biodiesels: a comparative study with diesel fuel. Sustainability 2021; 13(14): 7688.

19.

Shangjun

Chao

Yilin

, et al. Real-time model predictive control study of run-of-river hydropower plants with data-driven and physics-based coupled model. J Hydrol 2023; 617: 128942.

20.

Oney

Eric

OV.

Combining physics-based and data-driven modeling in well construction: hybrid fluid dynamics modeling. J Nat Gas Sci Eng 2022; 97: 104348.

21.

Liu

Liang

A physics-based and data-driven hybrid modeling method for accurately simulating complex contact phenomenon. Multibody Syst Dyn 2020; 50(1): 97–117.

22.

Wang

Xiao

Data-driven CFD modeling of turbulent flows through complex structures. Int J Heat Fluid Flow 2016; 62: 138–149.

23.

Stojanovic

Abdullah

Grujic

, et al. The influence of spoiler on the aerodynamic performances and longitudinal stability of the passenger car under high speed condition. J Vis 2022; 26(1): 97–112.

24.

Belhocine

Omar

WZW

. Three-dimensional finite element modeling and analysis of the mechanical behavior of dry contact slipping between the disc and the brake pads. Int J Adv Manuf Technol 2017; 88(1–4): 1035–1051.

25.

Shenghua

Yifan

Thomas

, et al. Integrating data-driven and physics-based approaches to characterize failures of interdependent infrastructures. Int J Crit Infrastruct Prot 2020; 31: 100391.

26.

Aleksandr

Joaquín

Measurements of user and sensor data from the internet of things (IoT) devices. Data 2022; 7(5): 52.

27.

Brahms

Benfano

Liquid level accuracy measurement using internet of things (IoT)-based multiple sensors in industrial mixing process. 革新的コンピューティング・情報・制御に関する速報 2022; 16(7): 769.

28.

Zubairi

Asri

Abd Malek

, et al. Usage of IOT and remote monitoring in a telemedicine application: a descriptive data. Int J Cardiol 2022; 369: 22.

29.

Noerhayati

Suprapto

Rahmawati

, et al. Improving wastewater quality system using the internet of things-based phytoremediation method. J Ecol Eng 2023; 24(3): 254–262.

30.

Linda

Hassina

Chahrazed

Towards a distributed SaaS management system in a multi-cloud environment. Cluster Comput 2022; 25(6): 4051–4071.

31.

Volkova

Krupenski

Ledvanov

, et al. Energy cascade connection of a low-temperature district heating network to the return line of a high-temperature district heating network. Energy 2020; 198: 117304.

32.

Ethan

Daniela

Jonathon

How convolutional neural network architecture biases learned opponency and color tuning. Neural Comput 2021; 33(4): 858–898.

33.

Guemo

Prodanovic

Militzer

Simulation of runout table cooling. Steel Res Int 2019; 90(4): 1800361.

34.

Yen

Robust multiobjective optimization via evolutionary algorithms. IEEE Trans Evol Comput 2019; 23(2): 316–330.

35.

Gaunce

Mandeville

, et al. DC arc flash: expanded incident energy equation for 125 V substation battery backup systems. IEEE Trans Ind Appl 2020; 57(2): 1183–1192.

36.

Jie

Anmin

Cheng

, et al. Rapid measurement of the fourth-order texture coefficient by laser ultrasonic surface acoustic waves based on a neural network expert system. Appl Opt 2019; 58(3): 626–634.

37.

Rabiej

Application of the artificial neural network for identification of polymers based on their X-ray diffraction curves. Comput Mater Sci 2021; 186: 110042.

38.

Dempe

Khamisov

Kochetov

A special three-level optimization problem. J Global Optim 2020; 76(3): 519–531.

Parameter optimisation of the ventilation system for an underground power space using a hybrid model

Abstract

Keywords

Introduction

Hybrid model

Physical model

Singular value decomposition

Date-driven model

Modelling procedures

Air out experimental methodology

Physical model and boundary conditions

Mesh generation

Governing equation, numerical model and boundary conditions

Solution settings and convergence criteria

Multivariate optimisation sampling of the subsurface power space ventilation system

Results and discussions

Analysis of simulation results

Analysis of the velocity field

Analysis of the temperature field

Results of physical model

Comparison of fluent and order-reduced models

Analysis of the prediction model

Results of the data-driven model

Conclusions

Footnotes

Acknowledgements

Declaration of conflicting interests

Funding

Research ethics

ORCID iD

References