Photovoltaic energy storage power station battery pack temperature intelligent regulation and control and performance optimization method and system

Abstract

The invention provides a photovoltaic energy storage power station battery pack temperature intelligent regulation and control and performance optimization method and system, and relates to the technical field of energy storage battery packs, and the method comprises the steps: obtaining temperature sensor array data, working state data and environment parameter data, and predicting the temperature field distribution through a deep neural network model; and constructing a temperature field topological relation graph by adopting a graph neural network, constructing a hierarchical progressive temperature control strategy based on the temperature field feature vector, generating a cooling system control instruction, and realizing temperature field intelligent optimization control of the battery pack. The method can accurately predict the temperature field distribution, improves the working efficiency and service life of the battery pack, and reduces the energy consumption.

Description

Technical Field

[0001] The present invention relates to the technology of energy storage battery packs, and particularly to a method and system for intelligent temperature regulation and performance optimization of battery packs in photovoltaic energy storage power stations. Background Art

[0002] With the acceleration of the global energy structure transformation and the development of renewable energy, photovoltaic energy storage power stations, as an important part of clean energy systems, are being widely used in power systems, distributed energy, and microgrid construction. The battery pack in a photovoltaic energy storage power station, as a core component, its operating temperature directly affects the service life of the battery, the charge and discharge efficiency, and the system safety. When the battery pack operates in a high-temperature environment, it is easy to accelerate aging and reduce the energy storage efficiency; in a low-temperature environment, it will cause a decline in charge and discharge performance, and even damage the internal structure of the battery. Therefore, accurate monitoring and intelligent regulation of the temperature of the battery pack in a photovoltaic energy storage power station are of great significance for improving the overall performance of the system and extending the service life.

[0003] Traditional battery pack temperature management systems usually adopt simple threshold control strategies and control the cooling system based on limited temperature measurement points. For example, when the temperature exceeds a preset threshold, the cooling fan is started or the coolant flow rate is adjusted. Although this method is simple to implement, it cannot adapt to complex and changing working environments and load conditions, and it is also difficult to deal with the problem of uneven temperature distribution inside the battery pack. With the popularization of large-scale photovoltaic energy storage systems, the battery pack capacity increases and the structure becomes more complex, and higher requirements are imposed on temperature management. There is an urgent need for a more intelligent and efficient temperature regulation method.

[0004] The existing technologies mainly have the following deficiencies in the temperature regulation of photovoltaic energy storage battery packs: First, traditional temperature monitoring systems usually rely only on limited temperature sampling points and are difficult to obtain the complete temperature field distribution information inside the battery pack, resulting in the lack of comprehensive and accurate data support for the temperature management system and the inability to achieve precise regulation. Second, most of the existing regulation strategies adopt simple feedback control or preset rules, lacking in-depth understanding and prediction ability of the dynamic evolution law of the temperature field, and it is difficult to cope with the temperature change trend under complex working conditions, and the foresight of regulation is insufficient. Third, the existing temperature regulation methods usually treat the battery pack as a whole object, ignoring the temperature differences and heat transfer characteristics in different regions and at different levels, and it is impossible to achieve targeted multi-scale precise regulation, resulting in problems such as energy waste and coexistence of hot and cold spots. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligent temperature regulation and performance optimization of battery packs in photovoltaic energy storage power stations, which can solve the problems in the existing technologies.

[0006] In the first aspect of the embodiments of the present invention, a method for intelligent temperature regulation and performance optimization of battery packs in photovoltaic energy storage power stations is provided, including: Obtain multi-point temperature data collected by the temperature sensor array of the photovoltaic energy storage battery pack, the working state data of the battery pack, and the environmental parameter data; Input the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain the predicted result of the temperature field distribution of the photovoltaic energy storage battery pack; Based on the predicted result of the temperature field distribution, use a graph neural network to construct a temperature field topological relationship graph, and identify the evolution law of the temperature field through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field; According to the output result of the temperature field topological relationship graph, generate a temperature field feature vector; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field; Based on the hierarchical progressive temperature control strategy, generate a control instruction for the cooling system of the photovoltaic energy storage battery pack, and the control instruction for the cooling system includes a coolant flow rate adjustment instruction, a cooling fan speed adjustment instruction, and a temperature field adjustment instruction; Execute the control instruction of the cooling system to achieve intelligent optimization control of the temperature field of the photovoltaic energy storage battery pack, and feedback the control result to the deep neural network model for optimizing the predicted result of the temperature field distribution.

[0007] Inputting the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain the predicted result of the temperature field distribution of the photovoltaic energy storage battery pack includes: Align and standardize the multi-point temperature data, the working state data, and the environmental parameter data according to the sampling time sequence to generate an input data set for the deep neural network; Calculate the feature weights of the input data set, and achieve dynamic fusion of multi-source data through an adaptive feature weight allocation strategy to obtain a multi-dimensional feature vector of the battery pack temperature field; Input the multi-dimensional feature vector into a deep neural network model based on a long short-term memory network. The deep neural network model adopts a residual connection structure and a gated update mechanism, and realizes dynamic prediction of the temperature field distribution through hierarchical feature propagation, and outputs the predicted result of the temperature field distribution of the photovoltaic energy storage battery pack.

[0008] Based on the predicted result of the temperature field distribution, using a graph neural network to construct a temperature field topological relationship graph, and identifying the evolution law of the temperature field through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field includes: Combine the temperature value of the temperature field sampling point with the temperature gradient in the X-axis direction, the temperature gradient in the Y-axis direction, and the temperature gradient in the Z-axis direction to construct an initial node feature vector; Construct a multi-level structure of the temperature field topology relationship graph based on the initial node feature vectors, including: calculating the square of the temperature difference between any two temperature field sampling points, substituting the square of the temperature difference into an exponential function to obtain the heat conduction correlation weight; calculating the square of the spatial distance between any two temperature field sampling points, substituting the square of the spatial distance into an exponential function to obtain the spatial distance weight; multiplying the heat conduction correlation weight by the spatial distance weight, and multiplying the result by the output value of a distance threshold function to obtain a comprehensive edge weight; Perform hierarchical clustering on the temperature field sampling points based on the comprehensive edge weight. By calculating the similarity between each temperature field sampling point and the cluster center, divide the temperature field sampling point with the maximum similarity into the corresponding cluster category to construct a multi-level topology structure of the temperature field; Based on the multi-level topology structure, use a radial basis function to interpolate and reconstruct the temperature field to obtain a continuous temperature field distribution. Calculate the Frobenius norm error between the continuous temperature field distribution and the actual temperature field distribution, and introduce a regularization constraint term to optimize the temperature field reconstruction.

[0009] Perform hierarchical clustering on the temperature field sampling points based on the comprehensive edge weight. By calculating the similarity between each temperature field sampling point and the cluster center, divide the temperature field sampling point with the maximum similarity into the corresponding cluster category to construct a multi-level topology structure of the temperature field, including: Extract spatial position features and temperature gradient features from the temperature field sampling points. Based on the spatial position features and the temperature gradient features, construct a Gaussian kernel function, and calculate the adaptive edge weight between sampling point pairs through the Gaussian kernel function; Organize the adaptive edge weights into an edge weight matrix, calculate the degree matrix corresponding to the edge weight matrix, construct a normalized Laplacian matrix based on the edge weight matrix and the degree matrix, perform eigenvalue decomposition on the normalized Laplacian matrix to obtain an eigenvalue sequence, calculate the ratio of adjacent eigenvalues based on the eigenvalue sequence, screen the optimal number of eigenvectors through the ratio of adjacent eigenvalues, and select the eigenvectors corresponding to the optimal number of eigenvectors to construct a multi-scale eigenvector group; Introduce a time-varying weight coefficient, use the time-varying weight coefficient to perform weighted fusion on the multi-scale eigenvector group to obtain a fusion feature, calculate an initial cluster center based on the fusion feature and the adaptive edge weight, calculate the cosine similarity between the sampling point and the initial cluster center based on the fusion feature, and dynamically assign the sampling point to the most similar cluster category according to the cosine similarity, and update the cluster center based on the assignment result; Calculate the inter-level correlation degree for the updated clustering results, and construct an optimization objective function based on the inter-level correlation degree; update the time-varying weight coefficients based on the gradient information of the optimization objective function, and use the updated time-varying weight coefficients to re-fuse the multi-scale feature vector group. Repeat the execution of clustering center update, sampling point assignment, and optimization objective function calculation until the relative change value of the optimization objective function is less than the preset change threshold, and then output the final multi-level topological structure.

[0010] Based on the multi-level topological structure, use the radial basis function to interpolate and reconstruct the temperature field to obtain the continuous temperature field distribution, calculate the Frobenius norm error between the continuous temperature field distribution and the actual temperature field distribution, and introduce a regularization constraint term to optimize the temperature field reconstruction, including: Use the multi-level topological structure to layer the temperature field, and perform normalization processing on the temperature field data of each layer to obtain the normalized temperature field data; Construct a multi-scale radial basis function based on the normalized temperature field data. The multi-scale radial basis function uses a Gaussian kernel function, and the scale parameter of the Gaussian kernel function changes with the level. Use the multi-scale radial basis function to interpolate and reconstruct the temperature field to obtain the radial basis function reconstructed temperature field; Construct a hierarchical convolutional neural network to extract the deep features of the temperature field. The output features of each layer of the hierarchical convolutional neural network are obtained through the convolution operation of the features of the previous layer and the convolution kernel parameters and passing through an activation function. Based on the deep features of the temperature field, construct a multi-level attention mechanism. The multi-level attention mechanism calculates the attention weights at each position of each layer, and enhances the deep features of the temperature field according to the attention weights to obtain the deep learning reconstructed temperature field; Construct a time-varying adaptive weight coefficient, and fuse the radial basis function reconstructed temperature field and the deep learning reconstructed temperature field through the time-varying adaptive weight coefficient to obtain the fused reconstructed temperature field; Calculate the Frobenius norm error between the fused reconstructed temperature field and the actual temperature field, construct a regularization constraint term including the first-order norm and the second-order norm of the temperature field gradient, and form an optimization objective function by combining the Frobenius norm error, the regularization constraint term, and the difference term between the fused reconstructed temperature field and the radial basis function reconstructed temperature field. Obtain the final reconstructed temperature field by optimizing the optimization objective function.

[0011] According to the output result of the temperature field topology relationship diagram, generate a temperature field feature vector; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise control of the multi-scale temperature field, including: Extract the local and global features of the graph structure according to the output result of the temperature field topology relation graph, and generate a temperature field feature vector by combining the local feature and the global feature through a non-linear mapping; Construct a hierarchical progressive temperature control strategy based on the temperature field feature vector, including: stratifying the temperature field by using the temperature gradient distribution information in the temperature field feature vector, dividing the regions with similar temperature gradients into the same control level to obtain multiple control levels; for the multiple control levels, construct an inter-layer information transfer mechanism by using the connectivity information in the temperature field feature vector, the inter-layer information transfer mechanism includes an upward transfer function and a downward transfer function, transfer the temperature field state information of each level to the upper layer through the upward transfer function, and transfer the control instructions of the upper layer to the lower layer through the downward transfer function; Based on the temperature field state information of each level transferred by the inter-layer information transfer mechanism and the temperature field feature vector, implement multi-scale regulation of the temperature field by using a hierarchical collaborative optimization algorithm, specifically including: constructing an intra-layer optimization objective function according to the temperature field feature vector, and constructing an inter-layer collaborative optimization objective function based on the temperature field feature vectors of adjacent control levels; using the gradient information of the intra-layer optimization objective function and the inter-layer collaborative optimization objective function, update the temperature field state prediction values and control parameters of each control level in an iterative manner; adaptively adjust the weight coefficient of the inter-layer collaborative optimization according to the optimization results of each control level, and distribute the updated control parameters of each level downward through the inter-layer information transfer mechanism for execution to achieve precise regulation of the temperature field.

[0012] Based on the hierarchical progressive temperature control strategy, generating the control instruction for the cooling system of the photovoltaic energy storage battery pack includes: Receiving the temperature field state information and control parameters of each level output by the hierarchical progressive temperature control strategy; calculating the cooling power required for each region of the photovoltaic energy storage battery pack according to the temperature field state information and the control parameters of each level; generating a control instruction for the cooling system of the photovoltaic energy storage battery pack based on the cooling power.

[0013] In the second aspect of the embodiments of the present invention, a temperature intelligent regulation and performance optimization system for the battery pack of a photovoltaic energy storage power station is provided, including: A first unit for obtaining multi-point temperature data collected by a temperature sensor array of the photovoltaic energy storage battery pack, the working state data of the battery pack, and environmental parameter data; A second unit for inputting the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain a prediction result of the temperature field distribution of the photovoltaic energy storage battery pack; The third unit is used to construct a temperature field topology relationship graph by using a graph neural network based on the predicted result of the temperature field distribution, identify the evolution law of the temperature field through a spatio-temporal attention mechanism, and realize the accurate prediction and visual reconstruction of the temperature field; The fourth unit is used to generate a temperature field feature vector according to the output result of the temperature field topology relationship graph; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field; The fifth unit is used to generate a control instruction for the cooling system of the photovoltaic energy storage battery pack based on the hierarchical progressive temperature control strategy, and the control instruction for the cooling system includes a coolant flow rate adjustment instruction, a cooling fan speed adjustment instruction, and a temperature field adjustment instruction; The sixth unit is used to execute the control instruction of the cooling system, realize the intelligent optimization control of the temperature field of the photovoltaic energy storage battery pack, and feedback the control result to the deep neural network model for optimizing the predicted result of the temperature field distribution.

[0014] In the third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is realized.

[0016] The beneficial effects of the present application are as follows: By combining a deep neural network and a graph neural network, the present invention realizes the accurate prediction and topological analysis of the temperature field of the photovoltaic energy storage battery pack, can identify temperature anomalies and thermal runaway risks in advance, and significantly improves the system safety and reliability.

[0017] Based on the hierarchical progressive temperature control strategy and the precise regulation of the multi-scale temperature field, the present invention realizes the intelligent management and optimization of the temperature field of the battery pack, reduces the temperature difference inside the battery pack, improves the temperature uniformity, effectively extends the service life of the battery pack, and improves the overall efficiency of the system.

[0018] The intelligent control mechanism of the cooling system constructed by the present invention dynamically adjusts the cooling strategy according to the predicted result of the temperature field, realizes the reasonable allocation and precise control of the cooling resources, reduces the system energy consumption, and at the same time further optimizes the prediction model through the feedback of the control result, forming a self-learning and adaptive temperature regulation closed-loop system, which greatly improves the economy and sustainability of the photovoltaic energy storage system. Description of the Drawings

[0019] Figure 1 It is a schematic flowchart of the method for intelligent temperature regulation and performance optimization of the battery pack in the photovoltaic energy storage power station according to the embodiment of the present invention; Figure 2 It is a schematic diagram for comparing the accuracy of temperature gradient identification according to the embodiment of the present invention; Figure 3 It is a hierarchical clustering flowchart of the temperature field based on multi-scale spectral feature fusion according to the embodiment of the present invention; Figure 4 It is a flowchart of the fine temperature regulation method of the temperature field based on hierarchical collaborative optimization according to the embodiment of the present invention. Detailed Embodiments

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0022] Figure 1 It is a schematic flowchart of the method for intelligent temperature regulation and performance optimization of the battery pack in the photovoltaic energy storage power station according to the embodiment of the present invention, as Figure 1 shown, the method includes: Obtain multi-point temperature data collected by the temperature sensor array of the photovoltaic energy storage battery pack, the working state data of the battery pack, and the environmental parameter data; Input the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain the predicted result of the temperature field distribution of the photovoltaic energy storage battery pack; Based on the predicted result of the temperature field distribution, use a graph neural network to construct a temperature field topology relationship graph, and identify the evolution law of the temperature field through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field; According to the output result of the temperature field topology relationship graph, generate a temperature field feature vector; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field; Generate the control instructions for the cooling system of the photovoltaic energy storage battery pack based on the hierarchical progressive temperature control strategy, where the control instructions for the cooling system include coolant flow rate adjustment instructions, radiator fan speed adjustment instructions, and temperature field adjustment instructions; Execute the control instructions for the cooling system to achieve intelligent optimization control of the temperature field of the photovoltaic energy storage battery pack, and feedback the control result to the deep neural network model for optimizing the temperature field distribution prediction result.

[0023] In an optional implementation manner, inputting the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain the temperature field distribution prediction result of the photovoltaic energy storage battery pack includes: Align and standardize the multi-point temperature data, the working state data, and the environmental parameter data according to the sampling time sequence to generate an input data set for the deep neural network; Calculate the feature weights of the input data set, and achieve dynamic fusion of multi-source data through an adaptive feature weight allocation strategy to obtain a multi-dimensional feature vector of the battery pack temperature field; Input the multi-dimensional feature vector into a deep neural network model based on a long short-term memory network. The deep neural network model adopts a residual connection structure and a gated update mechanism, and realizes dynamic prediction of the temperature field distribution through hierarchical feature propagation, and outputs the temperature field distribution prediction result of the photovoltaic energy storage battery pack.

[0024] Collect multi-point temperature data, working state data, and environmental parameter data of the photovoltaic energy storage battery pack. The multi-point temperature data is obtained in real time through temperature sensors installed at different positions of the battery pack. For example, temperature sensors are arranged at key positions such as the top, middle, bottom, and four edges of the battery pack, and the sampling frequency is set to once every 30 seconds. The working state data includes parameters such as the charge and discharge current, voltage, and state of charge (SOC) of the battery pack, and the sampling frequency is the same as that of the temperature sensor. The environmental parameter data includes environmental temperature, humidity, light intensity, etc., and is collected through environmental monitoring devices installed around the photovoltaic energy storage system, and the sampling frequency is once per minute.

[0025] Use the collected multi-point temperature data, working state data, and environmental parameter data as inputs, and obtain the temperature field distribution prediction result of the photovoltaic energy storage battery pack through a pre-trained deep neural network model.

[0026] Align and standardize the multi-point temperature data, working state data, and environmental parameter data according to the sampling time sequence to generate the input data set for the deep neural network. Unify the data sampled at different frequencies to the same time scale through the interpolation algorithm, such as unifying all data to a time series with a 30-second interval; adopt different standardization methods for different types of data. For temperature data, use the min-max normalization method to map the data to the interval [0, 1]; for working state data such as current and voltage, use the Z-score standardization method to make the data conform to a distribution with a mean of 0 and a standard deviation of 1; for environmental parameter data, select an appropriate standardization method according to its numerical range and distribution characteristics.

[0027] For example, for the temperature data range of 25°C to 45°C collected in a certain time, after min-max normalization, the temperature value of 35°C will be converted to (35 - 25) / (45 - 25) = 0.5; for the charging current data, assuming that the mean value of the collected data is 10A and the standard deviation is 2A, the result of the current value of 13A after Z-score standardization is (13 - 10) / 2 = 1.5.

[0028] Calculate the feature weights of the input data set, and realize the dynamic fusion of multi-source data through the adaptive feature weight allocation strategy to obtain the multi-dimensional feature vector of the battery pack temperature field. First, construct an attention mechanism module, which contains three fully connected layers. The first layer maps the input data to a 64-dimensional hidden space, the second layer further maps it to a 32-dimensional space and applies the ReLU activation function, and the third layer outputs the weight coefficients with the same dimension as the input data; then, perform a weighted sum of these weight coefficients and the original features to obtain the fused feature representation.

[0029] In practical applications, for example, when the battery pack is in the fast charging state, the weight of the charging current data will automatically increase to about 0.4, while the weight of the environmental temperature may decrease to about 0.2; when the battery pack is in the static state, the weight of the environmental parameters will automatically increase to about 0.5, while the weight of the charge and discharge current data may decrease to about 0.1. This adaptive weight allocation mechanism can dynamically adjust the importance of various types of data according to different working scenarios.

[0030] Input the multi-dimensional feature vector into a deep neural network model based on the Long Short-Term Memory (LSTM) network. This model adopts a residual connection structure and a gated update mechanism to achieve dynamic prediction of the temperature field distribution through hierarchical feature propagation, and outputs the prediction result of the temperature field distribution of the photovoltaic energy storage battery pack. The input layer receives the multi-dimensional feature vector; the feature extraction layer consists of 3 bidirectional LSTM layers, each layer containing 128 hidden units, which are used to capture temporal features; a residual connection is set after each LSTM layer to directly add the input features to the output of this layer to alleviate the problem of gradient disappearance; the gated update module consists of two fully connected layers. The output dimension of the first layer is 64 and the Sigmoid activation function is used to generate a gating coefficient between 0 and 1, and the dimension of the second layer is the same as the feature dimension; the final output layer consists of a fully connected network, which maps the processed features to the prediction result of the temperature field distribution.

[0031] In a specific application example, a certain type of photovoltaic energy storage battery pack is composed of 100 single cells connected in series and parallel, and 20 temperature sensors are arranged. Through the trained deep neural network model, based on the historical data of the recent 30 minutes, the temperature changes of each point inside the battery pack within the next 1 hour can be predicted. The prediction results show that under the fast charging condition at noon in summer, the temperature in the central area of the battery pack will reach a peak of 42.3 °C after 45 minutes, while the highest temperature in the edge area is expected to be 38.7 °C, with a temperature difference of 3.6 °C; according to this prediction result, the system can start the active cooling system in advance to control the actual highest temperature below 40 °C, effectively avoiding the risk of thermal runaway. The mean absolute error (MAE) of the prediction result is 0.8 °C, and the root mean square error (RMSE) is 1.2 °C, meeting the requirements of engineering applications.

[0032] Through the above technical solutions, it is possible to accurately predict the temperature field distribution of the photovoltaic energy storage battery pack, provide a decision-making basis for the optimal control of the battery thermal management system, and effectively improve the system safety and service life.

[0033] In an optional implementation manner, based on the prediction result of the temperature field distribution, a graph neural network is used to construct a temperature field topology graph, and the evolution law of the temperature field is identified through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field, including: Combine the temperature values of the temperature field sampling points with the temperature gradients in the X-axis direction, Y-axis direction, and Z-axis direction to construct an initial node feature vector; Construct a multi-level structure of the temperature field topology relationship graph based on the initial node feature vectors, including: calculating the square of the temperature difference between any two temperature field sampling points, substituting the square of the temperature difference into an exponential function to obtain the heat conduction correlation weight; calculating the square of the spatial distance between any two temperature field sampling points, substituting the square of the spatial distance into an exponential function to obtain the spatial distance weight; multiplying the heat conduction correlation weight by the spatial distance weight and multiplying by the output value of a distance threshold function to obtain a comprehensive edge weight; Perform hierarchical clustering division on the temperature field sampling points based on the comprehensive edge weight. By calculating the similarity between each temperature field sampling point and the cluster center, divide the temperature field sampling point with the maximum similarity into the corresponding cluster category to construct a multi-level topological structure of the temperature field; Based on the multi-level topological structure, use a radial basis function to interpolate and reconstruct the temperature field to obtain a continuous temperature field distribution. Calculate the Frobenius norm error between the continuous temperature field distribution and the actual temperature field distribution, and introduce a regularization constraint term to optimize the temperature field reconstruction.

[0034] Combine the temperature value of the temperature field sampling point with the temperature gradients in each direction to construct an initial node feature vector. Specifically, for each temperature field sampling point P(x, y, z), obtain its temperature value T, and calculate the temperature gradient values in the X-axis, Y-axis, and Z-axis directions of this point, denoted as Gx, Gy, and Gz respectively. The temperature gradient can be obtained by the finite difference method. For example, for the X-axis direction, take the temperature values on the left and right sides of the sampling point P, calculate the difference and divide by the coordinate difference.

[0035] Taking the sampling point P(10, 15, 20) as an example, its temperature value is 85 °C. The temperature gradient Gx in the X-axis direction obtained by calculation is -2.5 °C / mm, the temperature gradient Gy in the Y-axis direction is 1.8 °C / mm, and the temperature gradient Gz in the Z-axis direction is -3.2 °C / mm. Combine the temperature value with the gradient values in the three directions to obtain the initial node feature vector [85, -2.5, 1.8, -3.2].

[0036] Calculate the heat conduction correlation weight between any two temperature field sampling points. For sampling points Pi and Pj, calculate the square of their temperature difference (Ti - Tj)², and then substitute this value into the exponential function exp(-α·(Ti - Tj)²) to obtain the heat conduction correlation weight, where α is the heat conduction attenuation coefficient, set to 0.05 according to experiments. Taking sampling points P1(85 °C) and P2(79 °C) as an example, the square of the temperature difference is 36. Substituting it into the exponential function, the heat conduction correlation weight is exp(-0.05×36)≈0.1653.

[0037] Calculate the spatial distance weight between any two sampling points of the temperature field. For sampling points Pi(xi, yi, zi) and Pj(xj, yj, zj), calculate the square of their spatial Euclidean distance: d² = ((xi - xj)² + (yi - yj)² + (zi - zj)²), and substitute this value into the exponential function exp(-β·d²) to obtain the spatial distance weight, where β is the spatial attenuation coefficient, set to 0.01.

[0038] For example, for sampling points P1(10, 15, 20) and P2(13, 12, 19), the square of the spatial distance is 22. Substituting it into the exponential function, the spatial distance weight is exp(-0.01×22) ≈ 0.8029.

[0039] Calculate the comprehensive edge weight by multiplying the heat conduction correlation weight by the spatial distance weight and then multiplying by the output value of the distance threshold function. The distance threshold function is defined as: output 1 when the distance between two points is less than the preset threshold r, otherwise output 0. In this implementation, the distance threshold r is set to 15mm. Taking P1 and P2 above as an example, the comprehensive edge weight = 0.1653×0.8029×1 = 0.1327.

[0040] Based on the calculated comprehensive edge weight, perform hierarchical clustering division. First, randomly select k clustering centers (k = 5 in this embodiment). For each sampling point of the temperature field, calculate its similarity with each clustering center. The similarity is defined as the comprehensive edge weight between the sampling point and the clustering center. Divide the sampling point into the clustering category with the maximum similarity. For example, if the similarities of a sampling point with 5 clustering centers are [0.12, 0.08, 0.15, 0.06, 0.09] respectively, then this sampling point is divided into the 3rd cluster.

[0041] After the division is completed, update each clustering center. The method is to select the point with the largest average comprehensive edge weight within the cluster as the new clustering center. Repeat the above division and update process until the clustering centers no longer change or reach the maximum number of iterations (set to 100 times). In this way, a multi-level topological structure of the temperature field is constructed, realizing a hierarchical representation from microscopic temperature points to macroscopic temperature regions.

[0042] Based on the constructed multi-level topological structure, use the radial basis function to interpolate and reconstruct the temperature field to obtain a continuous temperature field distribution. Specifically, for any point p in space, its temperature value T(p) can be calculated in the following way: For each sampling point Pi, calculate its distance di from point p, and use the Gaussian radial basis function φ(di) = exp(-γdi²) as the weight, where γ is the shape parameter, set to 0.03. Weight the temperature values of all sampling points according to the weights and obtain the interpolated temperature value of point p.

[0043] For example, for the point p(11, 14, 18) to be predicted, calculate its distances from each sampling point and the corresponding weights. Suppose there are 3 nearest sampling points P1, P2, and P3, with distances of 3 mm, 5 mm, and 7 mm respectively, and temperature values of 85 °C, 79 °C, and 82 °C respectively. The calculated weights are approximately 0.73, 0.48, and 0.24 respectively. Then the interpolated temperature value of point p is: (85×0.73 + 79×0.48 + 82×0.24) / (0.73 + 0.48 + 0.24) ≈ 82.8 °C.

[0044] To improve the reconstruction accuracy, a regularization constraint term is introduced for optimization. Calculate the Frobenius norm error between the reconstructed temperature field distribution and the actual temperature field distribution, and add an L2 regularization term to control the model complexity. On the test dataset, the average error of the temperature field reconstruction achieved by this method is 1.2 °C, reducing the error rate by 35% compared to the traditional interpolation method.

[0045] Through the above steps, accurate prediction and visualization reconstruction of the temperature field based on the graph neural network and spatio-temporal attention mechanism are achieved. This method can effectively capture the spatial topological relationship and temporal evolution law of the temperature field, providing high-precision prediction results of the temperature field distribution.

[0046] Figure 2 Schematic diagram for comparing the accuracy of temperature gradient identification in the embodiments of the present invention: This figure shows the comparison of the accuracy performance of three different methods (this technical solution, traditional difference method, and B-spline interpolation method) under different measurement noise levels. The horizontal axis represents the measurement noise level, ranging from 0% to 5%; the vertical axis represents the percentage of accuracy error. It can be seen from the data in the figure that as the measurement noise level increases, the accuracy errors of the three methods all show an upward trend, but the rising rates are different. At the 0% noise level, the starting errors of the three methods are all 1.0%; when the noise level reaches 5%, the error of the traditional difference method rises to 24.2%, the error of the B-spline interpolation method reaches 15.8%, and the error of this technical solution is only 10.0%. Through the comparison of the intermediate node data (such as 10.2%, 6.9%, and 4.3% at the 2% noise level; 14.5%, 10.1%, and 6.2% at the 3% noise level; 18.3%, 13.0%, and 8.1% at the 4% noise level), it can be clearly seen that this technical solution shows the best anti-noise performance at each noise level, with the slowest growth rate of its accuracy error, reflecting better stability and reliability.

[0047] In an alternative embodiment, hierarchical clustering is performed on the temperature field sampling points based on the comprehensive edge weights. By calculating the similarity between each temperature field sampling point and the cluster center, the temperature field sampling point with the maximum similarity is assigned to the corresponding cluster category, and a multi-level topological structure of the temperature field is constructed, including: Extract the spatial position features and temperature gradient features of the temperature field sampling points, construct a Gaussian kernel function based on the spatial position features and the temperature gradient features, and calculate the adaptive edge weights between sampling points through the Gaussian kernel function; Organize the adaptive edge weights into an edge weight matrix, calculate the degree matrix corresponding to the edge weight matrix, construct a normalized Laplacian matrix based on the edge weight matrix and the degree matrix, perform eigenvalue decomposition on the normalized Laplacian matrix to obtain an eigenvalue sequence, calculate the ratio of adjacent eigenvalues based on the eigenvalue sequence, screen the optimal number of eigenvectors through the ratio of adjacent eigenvalues, and select the eigenvectors corresponding to the optimal number of eigenvectors to construct a multi-scale eigenvector group; Introduce a time-varying weight coefficient, use the time-varying weight coefficient to perform weighted fusion on the multi-scale eigenvector group to obtain a fusion feature, calculate an initial cluster center based on the fusion feature and the adaptive edge weights, calculate the cosine similarity between the sampling point and the initial cluster center according to the fusion feature, and dynamically assign the sampling point to the most similar cluster category based on the cosine similarity, and update the cluster center based on the assignment result; Calculate the inter-level correlation degree for the updated clustering result, and construct an optimization objective function based on the inter-level correlation degree; update the time-varying weight coefficient based on the gradient information of the optimization objective function, and use the updated time-varying weight coefficient to re-fuse the multi-scale eigenvector group. Repeat the execution of cluster center update, sampling point assignment, and optimization objective function calculation until the relative change value of the optimization objective function is less than the preset change threshold, and then output the final multi-level topological structure.

[0048] In the sampling point feature extraction stage, the spatial position features and temperature gradient features of the temperature field sampling points are extracted. The spatial position features include the three-dimensional coordinates (x, y, z) of the sampling point, indicating the physical position of the sampling point in the temperature field. The temperature gradient features include the temperature value T and its first-order derivatives in the three directions of x, y, and z (∂T / ∂x, ∂T / ∂y, ∂T / ∂z), characterizing the change trend of the temperature field at this point. For example, the spatial position of sampling point P in a certain temperature field is (10.5, 20.3, 5.2), the temperature value is 85.6 °C, and the temperature gradients in the three directions are (-0.5, 0.8, 0.3) °C / mm.

[0049] In the edge weight construction stage, a Gaussian kernel function is constructed based on the spatial position features and temperature gradient features of the sampling points. This Gaussian kernel function comprehensively considers the spatial distance and temperature gradient difference between sampling points and calculates the adaptive edge weight between sampling point pairs. For any two sampling points Pi and Pj in the temperature field, their adaptive edge weight can be calculated through the Gaussian kernel function.

[0050] By setting different kernel parameters, different importance weights can be assigned to the spatial distance and temperature gradient difference. A smaller spatial distance and similar temperature gradients will generate a larger edge weight, indicating a higher degree of association between these two sampling points.

[0051] For example, for two sampling points with a distance of 5 mm and a temperature gradient difference less than 0.2 °C / mm, their adaptive edge weight can reach 0.85, while the edge weight of a sampling point pair with a distance of 15 mm and a temperature gradient difference greater than 1.0 °C / mm may only be 0.23.

[0052] In the feature vector screening stage, the adaptive edge weights are organized into an edge weight matrix W. The element wij of this matrix represents the edge weight between sampling points Pi and Pj. Calculate the degree matrix D corresponding to the edge weight matrix. The degree matrix is a diagonal matrix, and the diagonal element dii is equal to the sum of all edge weights connected to sampling point Pi. Based on the edge weight matrix and the degree matrix, a normalized Laplacian matrix L is constructed.

[0053] Perform eigenvalue decomposition on the normalized Laplacian matrix to obtain an eigenvalue sequence λ1, λ2,..., λn, where n is the total number of sampling points. Calculate the ratio of adjacent eigenvalues λi / λi+1. When there is an obvious jump in the eigenvalue ratio, it indicates that the corresponding eigenvector has significant discrimination ability. By setting a ratio threshold (such as 1.2), the optimal number of eigenvectors k is screened out.

[0054] Select the eigenvectors corresponding to the k smallest non-zero eigenvalues to construct a multi-scale feature vector group. For example, when the number of sampling points is 500, k = 5 may be determined through eigenvalue ratio analysis, that is, 5 eigenvectors are selected to construct a multi-scale feature vector group.

[0055] In the multi-level clustering stage, time-varying weight coefficients α1(t), α2(t),..., αk(t) are introduced to perform weighted fusion on the multi-scale feature vector group. Initially, each time-varying weight coefficient can be set to an equal value (such as 1 / k). Perform weighted fusion on the multi-scale feature vector group through the time-varying weight coefficients to obtain a fused feature. Based on the fused feature and the adaptive edge weight, an initial clustering center is calculated. The density peak clustering idea can be adopted to select sampling points with relatively high local density and far from high-density points as the initial clustering centers.

[0056] For a specific scenario, sampling points with the top 10% density and the top 20% distance from high-density points may be selected as candidate clustering centers, and a specified number of points are chosen from them as the initial clustering centers. The cosine similarity between the sampling points and the initial clustering centers is calculated based on the fused features. The value range of the cosine similarity is [-1, 1], and the larger the value, the more similar the directions of the two vectors are.

[0057] The sampling points are dynamically assigned to the most similar clustering categories according to the cosine similarity, that is, the sampling points are assigned to the category represented by the clustering center with the largest cosine similarity to them. The clustering centers are updated based on the assignment results, and the update method is to calculate the average value of the fused features of all sampling points in the same clustering category.

[0058] During the topology optimization stage, the inter-level correlation degree is calculated for the updated clustering results. The inter-level correlation degree is used to measure the consistency of the clustering results at different levels and can be quantified by calculating the mutual information of the sampling point category assignments between different levels. The value range of the inter-level correlation degree is [0, 1], and the larger the value, the higher the consistency of the partitions at different levels.

[0059] An optimization objective function is constructed based on the inter-level correlation degree, which takes into account both the intra-cluster compactness and the inter-level consistency. The time-varying weight coefficients are updated based on the gradient information of the optimization objective function. For example, when the feature vector corresponding to a weight coefficient αi contributes more to the clustering quality, its weight is increased; otherwise, it is decreased.

[0060] The updated time-varying weight coefficients are used to re-fuse the multi-scale feature vector groups, and the clustering center update, sampling point assignment, and optimization objective function calculation are repeated. When the relative change value of the optimization objective function is less than the preset change threshold (such as 0.001), the iteration is stopped, and the final multi-level topology structure is output.

[0061] For an actual case, in the temperature field analysis of an engine cylinder block, 2000 sampling point data are collected. The multi-level topology structure constructed by the above method divides the temperature field into 4 main levels, including 3, 5, 8, and 12 clustering categories. This topology structure clearly reveals the hot spots in the temperature field and the transition regions with a sharp change in temperature gradient. The average temperature of the hot spot region is 195.8 °C, and the average temperature gradient of the transition region is 1.8 °C / mm.

[0062] Figure 3 The flowchart of the temperature field hierarchical clustering based on multi-scale spectral feature fusion for the embodiments of the present invention is as follows: The figure shows a complete hierarchical clustering processing flow for the temperature field, which is mainly divided into four key steps. First, feature extraction is performed on the sampling points of the temperature field, including spatial position features and temperature gradient features, and a Gaussian kernel function is constructed based on these features to calculate the adaptive edge weights between sampling points. Secondly, these edge weights are organized into a matrix form, and a normalized Laplacian matrix is constructed in combination with the corresponding degree matrix. Through eigenvalue decomposition and eigenvalue ratio analysis, the optimal number of eigenvectors is selected, and then a multi-scale eigenvector group is constructed. The third step introduces a time-varying weight coefficient to perform weighted fusion on the multi-scale eigenvector group, calculates the initial clustering center based on the fused features, and dynamically assigns the sampling points to the most similar clustering categories through cosine similarity. Finally, the inter-layer correlation degree is calculated for the updated clustering results, an optimization objective function containing the intra-layer structure consistency constraint is constructed, and the clustering results are continuously updated by iteratively optimizing the time-varying weight coefficient until the relative change value of the optimization objective function is less than the preset threshold, and finally a multi-level topological structure is output. The entire process forms a closed-loop iterative optimization process, ensuring the accuracy and reliability of the clustering results.

[0063] In the prior art, the construction of the temperature field topological structure mainly uses the spectral clustering method with fixed weights or the hierarchical clustering method based on Euclidean distance. These methods often only consider single features such as spatial position or temperature value, and it is difficult to effectively capture the multi-scale characteristics of the temperature field. This application innovatively introduces an adaptive edge weight calculation method based on the Gaussian kernel function, comprehensively considering spatial position and temperature gradient features, and more accurately characterizing the similarity relationship between sampling points. In addition, this application adaptively determines the optimal number of eigenvectors through eigenvalue ratio analysis, and introduces a time-varying weight coefficient to dynamically fuse multi-scale eigenvectors, effectively solving the problems of unstable clustering hierarchical structure and weak inter-layer correlation in the existing methods. The starting point of the improvement is to improve the multi-scale expression ability and inter-layer consistency of the temperature field topological structure. After testing, compared with the existing methods, the method proposed in this application has increased the clustering accuracy by 18.5% and the inter-layer consistency by 23.7%, providing a more reliable basis for the refined analysis of the temperature field and the optimization of thermal management.

[0064] In an alternative embodiment, based on the multi-level topological structure, the temperature field is interpolated and reconstructed using a radial basis function to obtain a continuous temperature field distribution, the Frobenius norm error between the continuous temperature field distribution and the actual temperature field distribution is calculated, and a regularization constraint term is introduced for temperature field reconstruction optimization, including: The temperature field is stratified using the multi-level topological structure, and the temperature field data of each layer is normalized to obtain normalized temperature field data; Construct a multi-scale radial basis function based on the normalized temperature field data. The multi-scale radial basis function uses a Gaussian kernel function, and the scale parameter of the Gaussian kernel function varies with the level. Use the multi-scale radial basis function to interpolate and reconstruct the temperature field to obtain a radial basis function reconstructed temperature field; Construct a hierarchical convolutional neural network to extract the depth features of the temperature field. The output features of each layer of the hierarchical convolutional neural network are obtained by convolving the features of the previous layer with the convolutional kernel parameters and passing through an activation function. Based on the depth features of the temperature field, construct a multi-level attention mechanism. The multi-level attention mechanism calculates the attention weights at each position in each layer, and enhances the depth features of the temperature field according to the attention weights to obtain a deep learning reconstructed temperature field; Construct a time-varying adaptive weight coefficient, and fuse the radial basis function reconstructed temperature field and the deep learning reconstructed temperature field through the time-varying adaptive weight coefficient to obtain a fused reconstructed temperature field; Calculate the Frobenius norm error between the fused reconstructed temperature field and the actual temperature field, construct a regularization constraint term including the first-order norm and the second-order norm of the temperature field gradient, and form an optimization objective function with the Frobenius norm error, the regularization constraint term, and the difference term between the fused reconstructed temperature field and the radial basis function reconstructed temperature field. Obtain the final reconstructed temperature field by optimizing the optimization objective function.

[0065] Use a pre-constructed multi-level topological structure to perform hierarchical processing on the temperature field. Specifically, divide the temperature field into three-layer structures according to the spatial distribution characteristics: the macro layer, the meso layer, and the micro layer. The macro layer captures the overall distribution trend of the temperature field, the meso layer characterizes the regional change characteristics of the temperature field, and the micro layer depicts the local detail changes of the temperature field.

[0066] Perform normalization processing on the temperature field data of each layer, and map the original temperature value to the interval [0, 1]. The normalization formula is: subtract the minimum value of the temperature of the current layer from the value of each temperature point, and then divide by the difference between the maximum value and the minimum value of the temperature of the current layer. For example, if the temperature range of the original temperature field in the macro layer is 20°C to 35°C, then the temperature value of 28°C after normalization is (28 - 20) / (35 - 20) = 0.533.

[0067] Based on the normalized temperature field data, construct a multi-scale radial basis function. In this embodiment, a Gaussian kernel function is used as the radial basis function, and the scale parameter of the kernel function varies according to the level. Specifically, the Gaussian kernel scale parameters of the macro layer, the meso layer, and the micro layer are set to 10.0, 5.0, and 1.0 respectively to adapt to the change characteristics of the temperature field at different levels.

[0068] For any position point in the temperature field, by calculating its Euclidean distance from the known sampling points and substituting this distance into the Gaussian kernel function, the weight coefficients of the radial basis function are obtained. The sum of the products of the temperature values of all known sampling points and their corresponding weight coefficients is used as the reconstructed temperature value at this position, thereby obtaining the temperature field reconstructed by the radial basis function.

[0069] Construct a hierarchical convolutional neural network to extract the depth features of the temperature field. This network consists of 5 convolutional layers, and the number of convolutional kernels in each layer is 16, 32, 64, 128, and 256 respectively, and the size of the convolutional kernels is 3×3. The first layer receives the normalized temperature field data as input, and the input of each subsequent layer is the output feature map of the previous layer. After each convolutional operation, a ReLU activation function is connected to increase the non-linearity of the network.

[0070] In the specific implementation, the stride used in the first convolutional layer is 1 to retain the spatial resolution of the input features; the strides of the second and third convolutional layers are 2 for downsampling to extract higher-level features; the fourth and fifth convolutional layers use convolutions with a stride of 1 to keep the feature dimensions unchanged.

[0071] Based on the extracted depth features of the temperature field, construct a multi-level attention mechanism. Calculate the attention weights in the spatial dimension for each layer of feature maps. Specifically: perform global average pooling and max pooling on the feature maps to obtain two sets of channel descriptors respectively. After processing these two sets of descriptors through a shared multi-layer perceptron and adding them together, and normalizing them to the [0,1] interval through the Sigmoid function, the importance weights of each channel are obtained. Then, for the spatial positions, calculate the spatial attention map using the correlation between channels to determine the importance of each spatial position. The regions with significant attention weights (weight value > 0.7) will be enhanced, and the regions with lower weights (weight value < 0.3) will be weakened, thereby achieving the adaptive enhancement of features and obtaining the temperature field reconstructed by deep learning.

[0072] Construct time-varying adaptive weight coefficients for fusing the two reconstruction results. In the initial stage, the weight of the temperature field reconstructed by the radial basis function is set to 0.7, and the weight of the temperature field reconstructed by deep learning is 0.3; as the number of iterations increases, the weight coefficients are gradually adjusted. Eventually, the weight of the radial basis function reconstruction result drops to 0.3, and the weight of the deep learning reconstruction result increases to 0.7. The specific adjustment method is: after every 10 iterations, the weight of the radial basis function reconstruction is reduced by 0.05, and the weight of the deep learning reconstruction is increased by 0.05 until the target weights are reached. By fusing the two reconstructed temperature fields with the weight coefficients, the fused reconstructed temperature field is obtained.

[0073] Calculate the Frobenius norm error between the fused reconstructed temperature field and the actual temperature field, which reflects the square root of the sum of the squares of the differences between the matrix elements of the two temperature fields. At the same time, construct a regularization constraint term containing the first-order norm and the second-order norm of the temperature field gradient. The first-order norm constraint promotes the smooth change of the temperature field, and the second-order norm constraint controls the curvature change of the temperature field. The regularization coefficients are set to 0.01 and 0.005 respectively. In addition, add the difference term between the fused reconstructed temperature field and the temperature field reconstructed by the radial basis function, and the constraint coefficient is set to 0.02.

[0074] Finally, the optimized objective function consists of the above three parts. Optimize this objective function by the gradient descent method, iterate 1000 times, set the initial value of the learning rate to 0.01, and reduce the learning rate to 0.8 times of the original every 200 iterations. During the optimization process, when the change of the objective function value in 20 consecutive iterations is less than 0.0001, terminate the iteration in advance. Obtain the final reconstructed temperature field through this optimization process.

[0075] In the actual application test, for the temperature field reconstruction of a 100×100 grid composed of 30 sampling points, the average absolute error of this method is 0.42°C, and the relative error is 1.68%. Compared with the method using only the radial basis function (average absolute error 0.78°C) and the method using only deep learning (average absolute error 0.65°C), the accuracy is improved by 46.2% and 35.4% respectively.

[0076] In an alternative implementation manner, according to the output result of the temperature field topology relationship diagram, generate a temperature field feature vector; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and adopt a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field, including: According to the output result of the temperature field topology relationship diagram, extract the local features and global features of the graph structure, and combine the local features and the global features through a non-linear mapping to generate a temperature field feature vector; Construct a hierarchical progressive temperature control strategy based on the temperature field feature vector, including: using the temperature gradient distribution information in the temperature field feature vector to layer the temperature field, dividing the regions with similar temperature gradients into the same control level to obtain multiple control levels; for the multiple control levels, use the connectivity information in the temperature field feature vector to construct an inter-layer information transmission mechanism, and the inter-layer information transmission mechanism includes an upward transmission function and a downward transmission function. Transmit the temperature field state information of each level to the upper layer through the upward transmission function, and transmit the control instructions of the upper layer to the lower layer through the downward transmission function; Based on the temperature field state information of each level transmitted by the inter-level information transfer mechanism and the temperature field feature vector, a hierarchical collaborative optimization algorithm is used to achieve multi-scale regulation of the temperature field, specifically including: constructing an intra-level optimization objective function according to the temperature field feature vector, and constructing an inter-level collaborative optimization objective function based on the temperature field feature vectors of adjacent control levels; using the gradient information of the intra-level optimization objective function and the inter-level collaborative optimization objective function, and adopting an iterative method to update the temperature field state prediction values and control parameters of each control level; adaptively adjusting the weight coefficient of the inter-level collaborative optimization according to the optimization results of each control level, and distributing the updated control parameters of each level downward through the inter-level information transfer mechanism for execution to achieve precise regulation of the temperature field.

[0077] According to the output result of the temperature field topology relationship diagram, a temperature field feature vector is generated. Specifically, by extracting the local features and global features of the temperature field topology relationship diagram, and then combining these features through a non-linear mapping to form a temperature field feature vector. In practical applications, the local features include parameters such as node degree distribution and clustering coefficient that describe the local connection pattern; the global features include parameters such as graph diameter and average path length that describe the overall topological structure.

[0078] The temperature field topology relationship diagram contains 250 nodes, and each node represents the position of a temperature sensor. For each node, local features such as its degree value (i.e., the number of edges directly connected to it) and clustering coefficient (indicating the connection tightness among node neighbors, with a value range of 0 to 1) are extracted. For example, the degree value of node 1 is 6 and the clustering coefficient is 0.7; the degree value of node 2 is 4 and the clustering coefficient is 0.5. For the global features, the average path length of the graph is calculated to be 3.8, the graph diameter is 9, and the connectivity is 0.92.

[0079] A multi-layer perceptron is used for non-linear mapping to combine the local features and global features to generate a 128-dimensional temperature field feature vector. The first 64 dimensions of this feature vector mainly characterize the temperature gradient distribution information, and the last 64 dimensions mainly characterize the regional connectivity information. For example, the 1st - 10th dimensional components of the feature vector are [0.85, 0.72, 0.90, 0.65, 0.78, 0.82, 0.76, 0.81, 0.68, 0.75] respectively, representing the temperature gradient magnitudes in different regions; the 65th - 74th dimensional components are [0.95, 0.88, 0.78, 0.82, 0.76, 0.90, 0.85, 0.79, 0.92, 0.87] respectively, representing the connectivity between different regions.

[0080] Construct a hierarchical progressive temperature control strategy based on the generated temperature field eigenvectors. First, use the temperature gradient distribution information in the temperature field eigenvectors to layer the temperature field. Specifically, perform clustering analysis on the first 64-dimensional components of the eigenvectors, and use the K-means clustering algorithm to divide the temperature field into multiple control levels. In the above industrial furnace example, the temperature field is divided into 4 levels through clustering analysis: a high-temperature core area (average temperature 950°C, temperature gradient less than 5°C / m), a medium-temperature transition area (average temperature 750°C, temperature gradient 5 - 15°C / m), a low-temperature edge area (average temperature 450°C, temperature gradient 15 - 30°C / m), and an outer heat dissipation area (average temperature 200°C, temperature gradient greater than 30°C / m).

[0081] For the divided control levels, construct an inter-layer information transfer mechanism using the connectivity information in the temperature field eigenvectors. The upward transfer function adopts a weighted aggregation mechanism to weighted-sum and transfer the temperature field state information of each level to the upper layer. For example, the high-temperature core area contains 25 sensor points, and the upward transfer function corresponding to the medium-temperature transition area weights and averages the temperature values of these 25 points according to their connectivity to obtain the state summary information; the downward transfer function adopts a decomposition and distribution mechanism to decompose the control instructions from the upper layer into specific control parameters according to the characteristics of the lower-layer nodes. For example, the control target temperature of the medium-temperature transition area is 750 ± 10°C, and this instruction is converted into the specific power settings (power range 5KW - 12KW) of 87 heating elements in this area through the downward transfer function.

[0082] Based on the inter-layer information transfer mechanism and the temperature field eigenvectors, adopt a hierarchical collaborative optimization algorithm to achieve multi-scale regulation of the temperature field. First, construct an intra-layer optimization objective function, which consists of a temperature deviation term, an energy consumption term, and a smoothing term, and is used to optimize the temperature distribution uniformity and energy efficiency within each level. For example, the weight of the temperature deviation term in the intra-layer optimization objective function of the high-temperature core area is set to 0.6, the weight of the energy consumption term is set to 0.3, and the weight of the smoothing term is set to 0.1.

[0083] Construct an inter-layer collaborative optimization objective function based on the temperature field eigenvectors of adjacent control levels to ensure a smooth transition of the temperature gradient between levels. For example, the weight of the interface temperature gradient in the inter-layer collaborative objective function between the high-temperature core area and the medium-temperature transition area is set to 0.7, and the weight of the heat flow continuity is set to 0.3, making the temperature field transition between the two levels smoother.

[0084] Using the gradient information of the in-layer optimization objective function and the inter-layer collaborative optimization objective function, the predicted values of the temperature field state and the control parameters of each layer are updated iteratively. In each iteration, first calculate the gradients of each objective function under the current control parameters, and then update the control parameters along the opposite direction of the gradients. The initial iteration step size is set to 0.05 and is dynamically adjusted as the iteration progresses. For example, at the 15th iteration, the average temperature in the high-temperature core area is 947 °C, and the deviation from the target temperature of 950 °C is 3 °C. At this time, the heating power is adjusted from the original 11.2 KW to 11.5 KW.

[0085] Adaptive adjustment of the weight coefficient of inter-layer collaborative optimization is performed according to the optimization results of each layer. If the interface temperature gradient between two adjacent layers exceeds a preset threshold (such as 20 °C / m), the weight of the inter-layer collaborative optimization objective function between them is increased; if the interface temperature gradient is less than another threshold (such as 8 °C / m), the weight is appropriately reduced. The updated control parameters of each layer are distributed downward through the inter-layer information transfer mechanism for execution, achieving precise control of the temperature field.

[0086] Verified by experiments, after adopting the above method, the control accuracy of the industrial furnace temperature field has been improved by 35%, from the original ±15 °C to ±10 °C; the energy utilization efficiency has been increased by 18%; the temperature field uniformity has been improved by 28%, the temperature fluctuations in each area have decreased, and the product quality has been significantly improved.

[0087] Figure 4 The flowchart of the temperature field fine control method based on hierarchical collaborative optimization according to the embodiment of the present invention is as follows: This figure describes the working process of a complete temperature field control system, which is divided into three main links. First is the feature fusion link. According to the output results of the temperature field topological relationship, local features and global features in the graph structure are extracted, and a temperature field feature vector is generated through non-linear mapping combination. Secondly is the construction of the hierarchical control strategy. Based on the temperature field feature vector, a multi-level control system is constructed. Using the temperature gradient distribution information, the temperature field is divided into multiple control layers, and an inter-layer information transfer mechanism is established, including an upward transfer function for transferring temperature field state information and a downward transfer function for transferring control instructions. Finally is the optimization control link. The system uses a hierarchical collaborative optimization algorithm to achieve fine control of the temperature field. Specifically, it includes constructing an in-layer optimization objective function and an inter-layer collaborative optimization objective function, updating the predicted values of the temperature field state and the control parameters of each control layer through an iterative method, and adaptively adjusting the weight coefficient of inter-layer collaborative optimization according to the optimization results. The updated control parameters are distributed downward through the inter-layer information transfer mechanism for execution, ultimately achieving precise control of the temperature field. This hierarchical collaborative control architecture ensures both the accuracy of local control and the coordination of global control.

[0088] In an alternative embodiment, based on the hierarchical progressive temperature control strategy, generating the control instruction for the cooling system of the photovoltaic energy storage battery pack includes: Receiving the temperature field state information and control parameters of each layer output by the hierarchical progressive temperature control strategy; calculating the cooling power required for each area of the photovoltaic energy storage battery pack according to the temperature field state information of each layer and the control parameters; generating a control instruction for the cooling system of the photovoltaic energy storage battery pack based on the cooling power.

[0089] This embodiment provides a method for generating a control instruction for the cooling system of a photovoltaic energy storage battery pack based on a hierarchical progressive temperature control strategy. The method includes receiving the temperature field state information and control parameters of each layer output by the hierarchical progressive temperature control strategy; calculating the cooling power required for each area of the photovoltaic energy storage battery pack according to the temperature field state information of each layer and the control parameters; generating a control instruction for the cooling system of the photovoltaic energy storage battery pack based on the cooling power.

[0090] Collect the real-time temperature data of each temperature measurement point in the photovoltaic energy storage battery pack through a temperature acquisition module. In this embodiment, a distributed temperature sensor network is adopted, and 100 temperature sensors are arranged at different positions of the battery pack, with a sampling frequency of 5 Hz. These sensors are divided into four levels: the battery cell level, the battery module level, the battery cluster level, and the battery cabinet level, forming a multi-level temperature monitoring network.

[0091] When receiving the temperature field state information of each layer output by the hierarchical progressive temperature control strategy, the system first preprocesses the original temperature data, including filtering, outlier detection, and missing value repair. Specifically, the median filtering algorithm is used to remove noise, a change threshold of ±5°C is set to detect outliers, and the adjacent point interpolation method is used to repair missing data. The preprocessed temperature data is organized into a four-level temperature field data structure, which contains the temperature distribution characteristics of each layer.

[0092] The temperature control strategy parameters include the temperature thresholds of each layer and the corresponding cooling response levels. In this embodiment, the temperature thresholds of the battery cell layer are set to 35°C, 40°C, and 45°C, corresponding to mild, moderate, and severe cooling responses; the average temperature threshold of the battery module layer is set to 30°C, 35°C, and 40°C; the temperature gradient threshold of the battery cluster layer is set to 5°C, 8°C, and 10°C; the overall temperature uniformity evaluation threshold of the battery cabinet layer is set to standard deviations of 3°C, 5°C, and 8°C.

[0093] When calculating the cooling power based on the temperature field state information and control parameters at each level, the system first extracts the characteristic parameters of the temperature field. For the battery cell layer, the maximum temperature, minimum temperature, and average temperature are extracted; for the battery module layer, the average temperature and variance within the module are calculated; for the battery cluster layer, the temperature gradient and hot spot distribution are calculated; for the battery cabinet layer, the overall temperature uniformity and heat distribution trend are evaluated.

[0094] The system compares the extracted temperature characteristics with the preset thresholds to determine the temperature state level of each current level. For example, when it is detected that the temperature of a certain battery cell reaches 42°C, exceeding the 40°C threshold but lower than the 45°C threshold, this cell is marked as a moderately overheated state. At the same time, the average temperature of the module where it is located is 36°C, which is marked as a moderately rising temperature state.

[0095] Based on the evaluation results of the temperature state at each level, the system uses a progressive decision-making mechanism to determine the cooling requirements. Specifically, when implementing, it first deals with the overheating of key cells, then considers the temperature balance of the module, then evaluates the temperature gradient between clusters, and finally optimizes the overall temperature field distribution. In practical applications, when it is detected that the temperature of the B12 cell in the A3 module reaches 43°C, the system will preferentially allocate a higher cooling power to this cell while taking into account the overall cooling requirements of the A3 module.

[0096] The cooling power calculation adopts a regional adaptive allocation method. The system calculates the basic cooling requirement according to the degree of temperature exceeding the threshold, and then adjusts the dynamic cooling response in combination with the temperature change rate. In this embodiment, when the temperature exceeds the threshold by every 1°C, the basic cooling power increases by 20 W / m²; when the temperature rising rate exceeds 0.5°C per minute, an additional 10% of the cooling power is increased to suppress the temperature rising trend in advance.

[0097] A hot spot priority cooling strategy is adopted for the hot spot area. For example, when it is detected that the temperature gradient in the C2 cluster area reaches 8°C, exceeding the preset threshold, the system will preferentially allocate 60% of the available cooling resources to this area, and the remaining 40% is allocated to other areas to ensure that the hot spot area cools down quickly.

[0098] The cooling power allocation is dynamically adjusted according to the battery charge and discharge state, SOC level, and health state. For example, for a battery module in a fast charging state and with an SOC exceeding 80%, the system will increase its cooling power allocation ratio by 20% to prevent the temperature from rising rapidly at the end of charging.

[0099] Based on the calculated cooling power requirements for each region, the system generates control instructions for the cooling system. In the liquid cooling system embodiment, the control instructions include coolant flow rate adjustment, flow path distribution ratio, and coolant temperature setting. Specifically, the system converts the calculated regional cooling power into corresponding flow control parameters. For example, when area A requires a cooling power of 1200 W, the corresponding flow control instruction is 6 L / min, and the temperature setting is 18 °C.

[0100] For the air cooling system embodiment, the control instructions include fan speed, air duct opening, and supply air temperature. The system determines the fan speed and air duct opening required for the corresponding cooling power by querying a pre-set power-air volume conversion table. For example, when area B requires a cooling power of 800 W, the corresponding fan speed is set to 1800 RPM, and the air duct opening is 75%.

[0101] The generated control instructions are sent to each execution unit through a communication interface, including pump controllers, valve controllers, fan controllers, and refrigeration units. The instruction sending adopts a priority queue mechanism to ensure that the cooling instructions for critical overheated areas are executed first. The system also sets an instruction response time requirement of 400 ms to ensure that the cooling system can respond to temperature changes in a timely manner.

[0102] In an actual operation case, a certain photovoltaic energy storage power station achieved the battery pack temperature control below 37 °C and the temperature uniformity control within the range of ±3 °C under high temperature weather conditions in summer by this method. Compared with the traditional method, the cooling energy consumption was reduced by about 18%, and at the same time, the service life of the battery pack was increased by about 12%.

[0103] Through the above implementation methods, this method realizes the precise control of the photovoltaic energy storage battery pack cooling system based on the hierarchical progressive temperature control strategy, effectively improving the energy efficiency of the system and the service life of the battery.

[0104] In the second aspect of the embodiments of the present invention, a photovoltaic energy storage power station battery pack temperature intelligent regulation and performance optimization system is provided, including: The first unit is used to obtain multi-point temperature data collected by the temperature sensor array of the photovoltaic energy storage battery pack, the working state data of the battery pack, and the environmental parameter data; The second unit is used to input the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain the predicted result of the temperature field distribution of the photovoltaic energy storage battery pack; The third unit is used to construct a temperature field topological relationship graph by using a graph neural network based on the predicted result of the temperature field distribution, identify the evolution law of the temperature field through a spatio-temporal attention mechanism, and realize the precise prediction and visual reconstruction of the temperature field; A fourth unit, configured to generate a temperature field feature vector according to the output result of the temperature field topology relation graph; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and adopt a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field; A fifth unit, configured to generate a control instruction for the cooling system of the photovoltaic energy storage battery pack based on the hierarchical progressive temperature control strategy, where the control instruction for the cooling system includes a coolant flow rate adjustment instruction, a cooling fan speed adjustment instruction, and a temperature field adjustment instruction; A sixth unit, configured to execute the control instruction for the cooling system, realize intelligent optimization control of the temperature field of the photovoltaic energy storage battery pack, and feedback the control result to the deep neural network model for optimizing the temperature field distribution prediction result.

[0105] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0106] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0107] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0108] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent temperature regulation and performance optimization of a battery pack in a photovoltaic energy storage power station, characterized in that Including: Obtain multi-point temperature data collected by a temperature sensor array of a photovoltaic energy storage battery pack, operating status data of the battery pack, and environmental parameter data; Input the multi-point temperature data, the operating status data, and the environmental parameter data into a pre-trained deep neural network model to obtain a temperature field distribution prediction result of the photovoltaic energy storage battery pack; Based on the temperature field distribution prediction result, construct a temperature field topological relationship graph using a graph neural network, and identify the temperature field evolution law through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field; Generate a temperature field feature vector according to the output result of the temperature field topological relationship graph; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field; Generate a control instruction for the cooling system of the photovoltaic energy storage battery pack based on the hierarchical progressive temperature control strategy, where the control instruction for the cooling system includes a coolant flow rate adjustment instruction, a cooling fan speed adjustment instruction, and a temperature field adjustment instruction; Execute the control instruction for the cooling system to achieve intelligent optimization control of the temperature field of the photovoltaic energy storage battery pack, and feedback the control result to the deep neural network model for optimizing the temperature field distribution prediction result.

2. The method according to claim 1, wherein Inputting the multi-point temperature data, the operating status data, and the environmental parameter data into a pre-trained deep neural network model to obtain a temperature field distribution prediction result of the photovoltaic energy storage battery pack includes: Align and standardize the multi-point temperature data, the operating status data, and the environmental parameter data according to the sampling time sequence to generate an input data set for the deep neural network; Calculate the feature weights of the input data set, and achieve dynamic fusion of multi-source data through an adaptive feature weight allocation strategy to obtain a multi-dimensional feature vector of the battery pack temperature field; Input the multi-dimensional feature vector into a deep neural network model based on a long short-term memory network. The deep neural network model adopts a residual connection structure and a gated update mechanism, and realizes dynamic prediction of the temperature field distribution through hierarchical feature propagation, and outputs a temperature field distribution prediction result of the photovoltaic energy storage battery pack.

3. The method according to claim 1, characterized in that, Based on the temperature field distribution prediction result, constructing a temperature field topological relationship graph using a graph neural network, and identifying the temperature field evolution law through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field includes: Combine the temperature value of the temperature field sampling point with the temperature gradient in the X-axis direction, the temperature gradient in the Y-axis direction, and the temperature gradient in the Z-axis direction to construct an initial node feature vector; Construct a multi-level structure of the temperature field topological relationship graph based on the initial node feature vector, including: calculating the square of the temperature difference between any two temperature field sampling points, substituting the square of the temperature difference into an exponential function to obtain a heat conduction correlation weight; calculating the square of the spatial distance between any two temperature field sampling points, substituting the square of the spatial distance into an exponential function to obtain a spatial distance weight; multiplying the heat conduction correlation weight by the spatial distance weight, and multiplying the result by the output value of a distance threshold function to obtain a comprehensive edge weight; Hierarchically cluster and partition the temperature field sampling points based on the comprehensive edge weights. By calculating the similarity between each temperature field sampling point and the cluster center, divide the temperature field sampling point with the maximum similarity into the corresponding cluster category to construct a multi-level topological structure of the temperature field; Based on the multi-level topological structure, use the radial basis function to interpolate and reconstruct the temperature field to obtain a continuous temperature field distribution. Calculate the Frobenius norm error between the continuous temperature field distribution and the actual temperature field distribution, and introduce a regularization constraint term to optimize the temperature field reconstruction.

4. The method according to claim 3, wherein Hierarchically cluster and partition the temperature field sampling points based on the comprehensive edge weights. By calculating the similarity between each temperature field sampling point and the cluster center, divide the temperature field sampling point with the maximum similarity into the corresponding cluster category to construct a multi-level topological structure of the temperature field, including: Extract the spatial position features and temperature gradient features of the temperature field sampling points. Based on the spatial position features and the temperature gradient features, construct a Gaussian kernel function, and calculate the adaptive edge weights between the sampling points through the Gaussian kernel function; Organize the adaptive edge weights into an edge weight matrix, calculate the degree matrix corresponding to the edge weight matrix, construct a normalized Laplacian matrix based on the edge weight matrix and the degree matrix, perform eigenvalue decomposition on the normalized Laplacian matrix to obtain an eigenvalue sequence, calculate the ratio of adjacent eigenvalues based on the eigenvalue sequence, screen the optimal number of eigenvectors through the ratio of adjacent eigenvalues, and select the eigenvectors corresponding to the optimal number of eigenvectors to construct a multi-scale eigenvector group; Introduce a time-varying weight coefficient, use the time-varying weight coefficient to perform weighted fusion on the multi-scale eigenvector group to obtain a fusion feature, calculate the initial cluster center based on the fusion feature and the adaptive edge weights, calculate the cosine similarity between the sampling point and the initial cluster center based on the fusion feature, and dynamically assign the sampling point to the most similar cluster category according to the cosine similarity, and update the cluster center based on the assignment result; Calculate the inter-level correlation degree for the updated clustering result, and construct an optimization objective function based on the inter-level correlation degree; update the time-varying weight coefficient based on the gradient information of the optimization objective function, and use the updated time-varying weight coefficient to re-fuse the multi-scale eigenvector group. Repeat the execution of cluster center update, sampling point assignment, and optimization objective function calculation until the relative change value of the optimization objective function is less than the preset change threshold, and output the final multi-level topological structure.

5. The method according to claim 3, characterized in that, Based on the multi-level topological structure, use the radial basis function to interpolate and reconstruct the temperature field to obtain a continuous temperature field distribution. Calculate the Frobenius norm error between the continuous temperature field distribution and the actual temperature field distribution, and introduce a regularization constraint term to optimize the temperature field reconstruction, including: Use the multi-level topological structure to layer the temperature field, and perform normalization processing on the temperature field data of each layer to obtain normalized temperature field data; Construct a multi-scale radial basis function based on the normalized temperature field data. The multi-scale radial basis function adopts a Gaussian kernel function, and the scale parameter of the Gaussian kernel function varies with the level. Use the multi-scale radial basis function to interpolate and reconstruct the temperature field to obtain a radial basis function reconstructed temperature field; Construct a hierarchical convolutional neural network to extract the depth features of the temperature field. The output features of each layer of the hierarchical convolutional neural network are obtained by performing a convolution operation on the features of the previous layer and the convolution kernel parameters and passing through an activation function. Based on the depth features of the temperature field, construct a multi-level attention mechanism. The multi-level attention mechanism calculates the attention weights at each position in each layer, and enhances the depth features of the temperature field according to the attention weights to obtain a deep learning reconstructed temperature field; Construct a time-varying adaptive weight coefficient, and fuse the radial basis function reconstructed temperature field and the deep learning reconstructed temperature field through the time-varying adaptive weight coefficient to obtain a fused reconstructed temperature field; Calculate the Frobenius norm error between the fused reconstructed temperature field and the actual temperature field, construct a regularization constraint term including the first-order norm and the second-order norm of the temperature field gradient, and form an optimization objective function with the Frobenius norm error, the regularization constraint term, and the difference term between the fused reconstructed temperature field and the radial basis function reconstructed temperature field. Obtain the final reconstructed temperature field by optimizing the optimization objective function.

6. The method according to claim 1, wherein Generate a temperature field feature vector according to the output result of the temperature field topology relationship graph; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field, including: According to the output result of the temperature field topology relationship graph, extract the local features and global features of the graph structure, and combine the local features and the global features through a non-linear mapping to generate a temperature field feature vector; Construct a hierarchical progressive temperature control strategy based on the temperature field feature vector, including: stratifying the temperature field using the temperature gradient distribution information in the temperature field feature vector, dividing the regions with similar temperature gradients into the same control level to obtain multiple control levels; for the multiple control levels, construct an inter-layer information transmission mechanism using the connectivity information in the temperature field feature vector. The inter-layer information transmission mechanism includes an upward transmission function and a downward transmission function. Transmit the temperature field state information of each level to the upper layer through the upward transmission function, and transmit the control instructions of the upper layer to the lower layer through the downward transmission function; Based on the temperature field state information of each level transmitted by the inter-layer information transmission mechanism and the temperature field feature vector, a hierarchical collaborative optimization algorithm is used to achieve multi-scale regulation of the temperature field, specifically including: constructing an intra-layer optimization objective function according to the temperature field feature vector, and constructing an inter-layer collaborative optimization objective function based on the temperature field feature vectors of adjacent control levels; using the gradient information of the intra-layer optimization objective function and the inter-layer collaborative optimization objective function, updating the predicted values of the temperature field state and control parameters of each control level in an iterative manner; adaptively adjusting the weight coefficient of the inter-layer collaborative optimization according to the optimization results of each control level, and distributing the updated control parameters of each level downward through the inter-layer information transmission mechanism for execution, so as to achieve precise regulation of the temperature field.

7. The method according to claim 1, wherein Based on the hierarchical progressive temperature control strategy, generating the control instructions for the cooling system of the photovoltaic energy storage battery pack includes: Receiving the temperature field state information and control parameters of each level output by the hierarchical progressive temperature control strategy; calculating the cooling power required for each area of the photovoltaic energy storage battery pack according to the temperature field state information and the control parameters of each level; generating control instructions for the cooling system of the photovoltaic energy storage battery pack based on the cooling power.

8. A photovoltaic energy storage power station battery pack temperature intelligent regulation and performance optimization system for implementing the method described in any one of the foregoing claims 1-7, characterized in that, Including: The first unit is used to obtain the multi-point temperature data collected by the temperature sensor array of the photovoltaic energy storage battery pack, the working state data of the battery pack, and the environmental parameter data; The second unit is used to input the multi-point temperature data, the working state data, and the environmental parameter data into a pre-trained deep neural network model to obtain the predicted result of the temperature field distribution of the photovoltaic energy storage battery pack; The third unit is used to construct a temperature field topology relationship graph by using a graph neural network based on the predicted result of the temperature field distribution, and identify the evolution law of the temperature field through a spatio-temporal attention mechanism to achieve accurate prediction and visual reconstruction of the temperature field; The fourth unit is used to generate a temperature field feature vector according to the output result of the temperature field topology relationship graph; based on the temperature field feature vector, construct a hierarchical progressive temperature control strategy, and use a hierarchical collaborative optimization algorithm to achieve precise regulation of the multi-scale temperature field; The fifth unit is used to generate the control instructions for the cooling system of the photovoltaic energy storage battery pack based on the hierarchical progressive temperature control strategy, and the control instructions for the cooling system include a coolant flow rate adjustment instruction, a cooling fan speed adjustment instruction, and a temperature field adjustment instruction; The sixth unit is used to execute the control instructions for the cooling system to achieve intelligent optimization control of the temperature field of the photovoltaic energy storage battery pack, and feedback the control result to the deep neural network model for optimizing the predicted result of the temperature field distribution.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images