Big data based electric pile energy consumption analysis

Abstract

The application provides a big data-based electric pile energy consumption analysis and prediction system, which extracts a variety of factors related to hydrogen energy consumption from massive data, such as in-vehicle state information, out-of-vehicle environment information perception and the like, further screens the related factors based on a reinforcement learning-based feature selection method, thereby mines important features affecting hydrogen fuel electric pile energy consumption, and realizes dynamic analysis on hydrogen fuel cell energy consumption in combination with a GRU model-based recurrent neural network, so as to predict the energy consumption trend of the electric pile, lay a foundation for subsequent hydrogen refueling demand calculation, ecological trajectory planning and battery energy management and the like, and has important significance for popularization and popularization of hydrogen fuel vehicles.

Description

Technical Field

[0001] This invention belongs to the field of hydrogen fuel cell stack energy consumption analysis technology, specifically relating to a system for analyzing and predicting the energy consumption of the stack based on big data from new energy vehicles. Background Technology

[0002] For pure electric vehicles, energy consumption is currently the most critical performance indicator of concern for electric vehicle drivers, automakers, and policymakers. Accurate and real-time prediction of energy consumption under real-world driving conditions is crucial for alleviating "range anxiety." However, most existing fuel cell stack energy consumption analyses are based on typical operating conditions, with very little comprehensive research on internal and external influencing factors, and experimental studies with small datasets are prone to random errors. The vast majority of existing technologies still focus on energy consumption analysis and prediction for lithium-ion batteries, with little attention paid to hydrogen fuel cells. For hydrogen fuel cells, a high-efficiency, low-pollution energy source representing the future direction of new energy development, how to provide a targeted analysis and prediction method suitable for hydrogen fuel cell stack energy consumption is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0003] To address the aforementioned technical problems in this field, the present invention provides a big data-based fuel cell stack energy consumption analysis and prediction system, specifically comprising the following components:

[0004] Data acquisition and processing module, feature selection module, energy consumption prediction model module;

[0005] The data acquisition and processing module is used to monitor the status of hydrogen fuel cell vehicles and to collect and store characteristic data related to fuel cell stack energy consumption. The data includes vehicle-related characteristic data, environmental-related characteristic data, and driver-related characteristic data. Preprocessing is performed on each relevant characteristic data to remove invalid data and fill missing data using a clustering-then-interpolation method. For the continuous operating conditions corresponding to the preprocessed relevant characteristic data, training sets, validation sets, and test sets are extracted and constructed respectively.

[0006] The feature selection module is used to select a feature set consisting of several features from all relevant features in the training set according to the correlation between each feature and the energy consumption of the fuel cell stack, so as to reduce the model training difficulty of the energy consumption prediction model module and improve the energy consumption prediction accuracy.

[0007] The energy consumption prediction model module uses the feature set to train the fuel cell stack energy consumption prediction model based on recurrent neural networks; after the model is trained, the model parameters are adjusted and optimized using the validation set, and the feature selection module is guided to update the selected feature set; the performance of the model is tested using the test set.

[0008] Furthermore, the vehicle-related feature data collected by the data acquisition and processing module specifically includes: power mode (DM). t The real-time voltage (V) of a single cell t ), temperature (T) t ), current (C) t ), vehicle speed (v) t ), vehicle acceleration (a t ), Real-time voltage of the regenerative braking system (BERS_V) t ), current (BERS_C) t ), power (BERS_P) t ), Power load operating status (PL) t ), hydrogen fuel system temperature (HS_T) t Hydrogen pressure (HS_P) t ), concentration (HS_C) t ), vehicle weight (Q t Hydrogen fuel consumption rate (H_rate) t The environmental-related characteristic data specifically includes: ambient temperature (DM). t Air resistance (DM) t ), slope resistance (DM) t Traffic conditions (DM) t Road conditions (DM) t The driver-related characteristic data specifically includes: vehicle operating mode (OM); t ), gear (Gx) t ), hydrogenation habit (HH) t Route planning (RP) t );

[0009] The power mode uses discrete coding, including: 0-pure hydrogen fuel cell vehicle, 1-hydrogen-electric hybrid vehicle, 2-vehicle with energy recovery function; the vehicle operation mode uses discrete coding, including: 0-normal mode, 1-economy mode, 3-sport mode; the ambient temperature (DM) t Air resistance (DM) t ), slope resistance (DM) t Traffic conditions (DM) t Road conditions (DM) t Data such as ) are collected from Baidu API and converted.

[0010] Furthermore, the preprocessing of the relevant feature data specifically includes:

[0011] First, remove vehicle data with no driving records or mileage exceeding 800km from the daily vehicle data. The log data is then segmented using a sliding window, with a time window of half an hour and a sliding step of 10 minutes, resulting in 143 operating condition segments for each vehicle per day. Vehicle data with a missing feature rate greater than 10% in any operating condition segment is removed. Then, for a specific vehicle's historical data, its operating condition segments are divided into 1000 classes using a clustering algorithm and missing features. Missing values ​​are repaired using interpolation with complete operating condition segments within the same class. The formula used is: The eigenvalues ​​are normalized to the range [0,1], where X is the original eigenvalue. max X is the maximum value of this eigenvalue. min This is the minimum value of the feature; the characteristic dynamic mode (DM) t ), vehicle operating mode (OM) t ), route planning (RP) t The road type and other type features are encoded using one-hot encoding.

[0012] Furthermore, the feature selection module selects a subset of features based on a reinforcement learning algorithm, specifically including the following steps:

[0013] 1) Calculate the information entropy of each relevant feature. Where p(x) i The characteristic value is equal to x. i The probability of; and, calculating the Pearson correlation coefficient between each pair of features. Where cov(X, Y) is the covariance between eigenvectors X and Y, and σ X σ Y Let X and Y be the standard deviations, respectively.

[0014] 2) Randomly initialize the array F = {f} of the current selected feature subset scheme. DM f V f T , ..., f RP The array dimension is equal to the number of features, and each array element is encoded using 0-1, where 0 indicates that the feature is not selected and 1 indicates that the feature is selected.

[0015] 3) Let the current state be S t Select the feature element f with the largest information entropy from the unselected feature subset. in Execute action a t Change the code 0 to 1; select the feature element f from the selected feature subset that has the largest sum of Pearson correlation coefficients with other features. out Execute action a t Change the encoding from 1 to 0, and set the feature selection state to S after the action is completed. t+1Calculate the benefits of the above actions using the following formula, and select the action that brings the greatest benefit:

[0016] Q(S t a t )=(1-ε)Q(S t a t )+ε(R(S t a t )+γ(maxQ(S t+1 a t+1 )))

[0017] Where ε (0 < ε < 1) is the learning rate controlling the convergence of the reinforcement learning algorithm, γ (0 ≤ γ ≤ 1) is the discount coefficient, and R(S t a t Q(S) represents the benefit gained from selecting or removing a feature. t a t The expression represents the immediate reward after selecting and executing an action, compared to the value obtained by executing the optimal strategy in subsequent cycles; that is, a comparison of the increased feature f. in With deletion feature f out For returns, select the action that yields the maximum cumulative return.

[0018] 4) Set up multiple agents and repeat the above steps until the maximum reward Q-value table stops updating, then stop iterating to obtain the selected feature set X.

[0019] Furthermore, the stack energy consumption prediction model specifically employs a recurrent neural network based on the GRU model, and the corresponding instantaneous benefit R(S) t a t The calculation method is as follows: First, construct a GRU model, train the model using the feature set selected each time, and use the feature vector of the test set corresponding to the feature set selected each time to make predictions, and calculate the deviation (mean square error, MSE) between the predicted hydrogen consumption and the actual hydrogen consumption.

[0020]

[0021] Where m is the number of samples in the test set, and H_rate is the set of hydrogen consumption rates within half an hour. These are the predicted values ​​from the GRU model.

[0022] Furthermore, using the selected feature set X, the GRU model performs the following computational process:

[0023] r t =σ(W r ·[h t-1 x t ])

[0024] z t =σ(W z ·[h t-1 x t ])

[0025]

[0026]

[0027]

[0028] Where, x t The input data at time t is taken from feature X, r t , z t These are the logic units for the update gate and the reset gate, respectively. t Controlling the network to discard or add information from previous time series, z t The control network determines the extent to which information is updated; h t-1 Let W be the hidden layer parameters of the network at time t-1. r W z , W o These are the weight matrices for the update gate, reset gate, state unit, and output layer, respectively. Let be the predicted hydrogen fuel consumption at time t, σ be the Sigmoid activation function, and tanh be the tanh activation function. They are calculated as follows:

[0029]

[0030]

[0031] A loss function is defined for the GRU model. The model parameters W are updated iteratively using backpropagation, with the goal of minimizing the model loss function. r W z , W o This continues until the model converges; the loss function of the model is designed as follows:

[0032]

[0033] During model training, the data in the feature set is first sorted by history, and then it is traversed step by step with multiple continuous working condition time windows as the length to obtain a series of continuous working condition segments. The last working condition data of the segment is used as the predicted value, and the first few working condition data of the segment are used as the feature values.

[0034] Furthermore, the energy consumption prediction model module specifically uses the PyTorch deep learning framework for model training, verification, and testing. By preprocessing the test set data using the same methods, standardized data input is obtained; the trained model file is loaded and run to obtain the predicted value of the test data; and the MSE between the predicted value and the true value is obtained according to the loss function calculation formula. The larger this value, the better the generalization ability of the prediction model.

[0035] The fuel cell stack energy consumption analysis and prediction system provided by the present invention extracts various factors related to hydrogen energy consumption from massive amounts of data, such as in-vehicle status information and external environmental information perception. Based on a reinforcement learning feature selection method, these related factors are further filtered to uncover important features affecting the energy consumption of hydrogen fuel cell stacks. Combined with a recurrent neural network based on the GRU model, the system achieves dynamic analysis of the energy consumption of hydrogen fuel cells and predicts the energy consumption trend of the stack. This lays the foundation for subsequent hydrogen refueling demand calculation, ecological trajectory planning, and battery energy management, and is of great significance for the promotion and popularization of hydrogen fuel cell vehicles. Attached Figure Description

[0036] Figure 1 The framework and workflow of the fuel cell stack energy consumption analysis and prediction system provided by this invention Detailed Implementation

[0037] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] The present invention provides a fuel cell stack energy consumption analysis and prediction system based on big data, such as... Figure 1 As shown, it specifically includes the following components:

[0039] Data acquisition and processing module, feature selection module, energy consumption prediction model module;

[0040] The data acquisition and processing module is used to monitor the status of hydrogen fuel cell vehicles and to collect and store characteristic data related to fuel cell stack energy consumption. The data includes vehicle-related characteristic data, environmental-related characteristic data, and driver-related characteristic data. Preprocessing is performed on each relevant characteristic data to remove invalid data and fill missing data using a clustering-then-interpolation method. For the continuous operating conditions corresponding to the preprocessed relevant characteristic data, training sets, validation sets, and test sets are extracted and constructed respectively.

[0041] The feature selection module is used to select a feature set consisting of several features from all relevant features in the training set according to the correlation between each feature and the energy consumption of the fuel cell stack, so as to reduce the model training difficulty of the energy consumption prediction model module and improve the energy consumption prediction accuracy.

[0042] The energy consumption prediction model module uses the feature set to train the fuel cell stack energy consumption prediction model based on recurrent neural networks; after the model is trained, the model parameters are adjusted and optimized using the validation set, and the feature selection module is guided to update the selected feature set; the performance of the model is tested using the test set.

[0043] In a preferred embodiment of the present invention, the vehicle-related feature data collected by the data acquisition and processing module specifically includes: power mode (DM). t The real-time voltage (V) of a single cell t ), temperature (T) t ), current (C) t ), vehicle speed (v) t ), vehicle acceleration (a t ), Real-time voltage of the regenerative braking system (BERS_V) t ), current (BERS_C) t ), power (BERS_P) t ), Power load operating status (PL) t ), hydrogen fuel system temperature (HS_T) t Hydrogen pressure (HS_P) t ), concentration (HS_C) t ), vehicle weight (Q t Hydrogen fuel consumption rate (H_rate) t The environmental-related characteristic data specifically includes: ambient temperature (DM). t Air resistance (DM) t ), slope resistance (DM) t Traffic conditions (DM) t Road conditions (DM) t The driver-related characteristic data specifically includes: vehicle operating mode (OM); t ), gear (Gx) t ), hydrogenation habit (HH) t Route planning (RP) t );

[0044] The power mode uses discrete coding, including: 0-pure hydrogen fuel cell vehicle, 1-hydrogen-electric hybrid vehicle, 2-vehicle with energy recovery function; the vehicle operation mode uses discrete coding, including: 0-normal mode, 1-economy mode, 3-sport mode; the ambient temperature (DM)t Air resistance (DM) t ), slope resistance (DM) t Traffic conditions (DM) t Road conditions (DM) t Data such as ) are collected from Baidu API and converted.

[0045] In a preferred embodiment of the present invention, the preprocessing of the relevant feature data specifically includes:

[0046] First, remove vehicle data with no driving records or mileage exceeding 800km from the daily vehicle data. The log data is then segmented using a sliding window, with a time window of half an hour and a sliding step of 10 minutes, resulting in 143 operating condition segments for each vehicle per day. Vehicle data with a missing feature rate greater than 10% in any operating condition segment is removed. Then, for a specific vehicle's historical data, its operating condition segments are divided into 1000 classes using a clustering algorithm and missing features. Missing values ​​are repaired using interpolation with complete operating condition segments within the same class. The formula used is: The eigenvalues ​​are normalized to the range [0,1], where X is the original eigenvalue. max X is the maximum value of this eigenvalue. min This is the minimum value of the feature; the characteristic dynamic mode (DM) t ), vehicle operating mode (OM) t ), route planning (RP) t The road type and other type features are encoded using one-hot encoding.

[0047] In a preferred embodiment of the present invention, the feature selection module selects a subset of features based on a reinforcement learning algorithm, specifically including the following steps:

[0048] 1) Calculate the information entropy of each relevant feature. Where p(x) i The characteristic value is equal to x. i The probability of; and, calculating the Pearson correlation coefficient between each pair of features. Where cov(X, Y) is the covariance between eigenvectors X and Y, and σ X σ Y Let X and Y be the standard deviations, respectively.

[0049] 2) Randomly initialize the array F = {f} of the current selected feature subset scheme. DM f V f T , ..., f RP The array dimension is equal to the number of features, and each array element is encoded using 0-1, where 0 indicates that the feature is not selected and 1 indicates that the feature is selected.

[0050] 3) Let the current state be S t Select the feature element f with the largest information entropy from the unselected feature subset. in Execute action a t Change the code 0 to 1; select the feature element f from the selected feature subset that has the largest sum of Pearson correlation coefficients with other features. out Execute action a t Change the encoding from 1 to 0, and set the feature selection state to S after the action is completed. t+1 Calculate the benefits of the above actions using the following formula, and select the action that brings the greatest benefit:

[0051] Q(S t a t )=(1-ε)Q(S t a t )+ε(R(S t a t )+γ(maxQ(S t+1 a t+1 )))

[0052] Where ε (0 < ε < 1) is the learning rate controlling the convergence of the reinforcement learning algorithm, γ (0 ≤ γ ≤ 1) is the discount coefficient, and R(S t a t Q(S) represents the benefit gained from selecting or removing a feature. t a t The expression represents the immediate reward after selecting and executing an action, compared to the value obtained by executing the optimal strategy in subsequent cycles; that is, a comparison of the increased feature f. in With deletion feature f out For returns, select the action that yields the maximum cumulative return.

[0053] 4) Set up multiple agents and repeat the above steps until the maximum reward Q-value table stops updating, then stop iterating to obtain the selected feature set X.

[0054] In a preferred embodiment of the present invention, the fuel cell stack energy consumption prediction model specifically employs a recurrent neural network based on the GRU model, and the corresponding instantaneous benefit R(S) t a t The calculation method is as follows: First, construct a GRU model, train the model using the feature set selected each time, and use the feature vector of the test set corresponding to the feature set selected each time to make predictions, and calculate the deviation (mean square error, MSE) between the predicted hydrogen consumption and the actual hydrogen consumption.

[0055]

[0056] Where m is the number of samples in the test set, and H_rate is the set of hydrogen consumption rates within half an hour. These are the predicted values ​​from the GRU model.

[0057] In a preferred embodiment of the present invention, the GRU model performs the following calculation process using the selected feature set X:

[0058] r t =σ(W r ·[h t-1 x t ])

[0059] z t =σ(W z ·[h t-1 x t ])

[0060]

[0061]

[0062]

[0063] Where, x t The input data at time t is taken from feature X, r t , z t These are the logic units for the update gate and the reset gate, respectively. t Controlling the network to discard or add information from previous time series, z t The control network determines the extent to which information is updated; h t-1 Let W be the hidden layer parameters of the network at time t-1. r W z , W o These are the weight matrices for the update gate, reset gate, state unit, and output layer, respectively. Let be the predicted hydrogen fuel consumption at time t, σ be the Sigmoid activation function, and tanh be the tanh activation function. They are calculated as follows:

[0064]

[0065]

[0066] A loss function is defined for the GRU model. The model parameters W are updated iteratively using backpropagation, with the goal of minimizing the model loss function. r W z , W oThis continues until the model converges; the loss function of the model is designed as follows:

[0067]

[0068] During model training, the data in the training set is first sorted by history, and then it is traversed step by step with a length of 8 consecutive working condition time windows to obtain a series of continuous working condition segments. The first 7 working condition data of the segment are used as feature values, and the last working condition data of the segment is used as the predicted value.

[0069] In a preferred embodiment of the present invention, the energy consumption prediction model module specifically uses the PyTorch deep learning framework for model training, verification, and testing. By preprocessing the test set data using the same data preprocessing methods, standardized data input is obtained; the trained model file is loaded and run to obtain the predicted value of the test data; and the MSE between the predicted value and the true value is obtained according to the loss function calculation formula. The larger the value, the better the generalization ability of the prediction model.

[0070] The aforementioned training data consists of 7 operating condition segments, with the prediction value being one future operating condition segment. The established operating condition segment duration is 30 minutes. The vehicle terminal uploads data at a frequency of one frame every 5 seconds, resulting in a data quantity of 30 × 7 × 60 ÷ 5 = 2520. The input sequence x at time t... t It can be represented in matrix form

[0071]

[0072] The future operating condition prediction sequence y is represented as:

[0073]

[0074] For the GRU model, the feature dimension of the hidden layer is set to 20, the number of training batches is 200, the optimization algorithm is Adam, and the number of iterations is set to 200.

[0075] It should be understood that the sequence number of each step in the embodiments of the present invention does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fuel cell stack energy consumption analysis and prediction system based on big data, characterized in that: Specifically, it includes the following components: Data acquisition and processing module, feature selection module, energy consumption prediction model module; The data acquisition and processing module is used to monitor the status of hydrogen fuel cell vehicles and to collect and store characteristic data related to fuel cell stack energy consumption. The data includes vehicle-related characteristic data, environmental-related characteristic data, and driver-related characteristic data. Preprocessing is performed on each relevant characteristic data to remove invalid data and fill missing data using a clustering-then-interpolation method. For the continuous operating conditions corresponding to the preprocessed relevant characteristic data, training sets, validation sets, and test sets are extracted and constructed respectively. The feature selection module is used to select a feature set consisting of several features from all relevant features in the training set based on the correlation between each feature and the energy consumption of the fuel cell stack. This reduces the model training difficulty of the energy consumption prediction model module and improves the accuracy of energy consumption prediction. Specifically, it selects a feature subset based on a reinforcement learning algorithm, including the following steps: 1) Calculate the information entropy of each relevant feature. ,in The characteristic value is equal to The probability of; and, calculating the Pearson correlation coefficient between each pair of features. ,in Let X be the covariance between the eigenvectors X and Y. , Let X and Y be the standard deviations, respectively. 2) Randomly initialize the current feature subset selection scheme array F. The array dimension is equal to the number of features. Each array element is encoded using 0-1, where 0 indicates that the feature is not selected and 1 indicates that the feature is selected. 3) Let the current state be Select the feature element with the highest information entropy from the unselected feature subset. Execute actions Change the code 0 to 1; select the feature element from the selected feature subset that has the largest sum of Pearson correlation coefficients with other features. Execute actions Change the encoding from 1 to 0, and set the feature selection state after the action is completed. Calculate the benefits of the above actions using the following formula, and select the action that brings the greatest benefit: in, It is the learning rate that controls the convergence of the reinforcement learning algorithm. This is the discount factor. The benefit derived from selecting or removing features at the current time. This represents the immediate reward after selecting and executing an action, compared to the value obtained by executing the optimal strategy in subsequent cycles; in other words, it's a comparison of added features. With deletion features For returns, select the action that yields the maximum cumulative return. 4) Set up multiple agents and repeat the above steps until the maximum reward Q-value table stops updating, at which point the iteration stops, and the selected feature set is obtained. ; The energy consumption prediction model module uses the feature set to train the fuel cell stack energy consumption prediction model based on recurrent neural networks; after the model is trained, the model parameters are adjusted and optimized using the validation set, and the feature selection module is guided to update the selected feature set; the performance of the model is tested using the test set.

2. The system as described in claim 1, characterized in that: The vehicle-related characteristic data collected by the data acquisition and processing module specifically includes: power mode, real-time voltage, temperature, and current of individual battery cells, vehicle speed, vehicle acceleration, real-time voltage, current, and power of the regenerative braking system, electrical load operating status, hydrogen fuel system temperature, hydrogen pressure, and concentration, vehicle weight, and hydrogen fuel consumption rate; the environmental-related characteristic data specifically includes: ambient temperature, air resistance, gradient resistance, traffic conditions, and road conditions; the driver-related characteristic data specifically includes: vehicle operating mode, gear selection, hydrogen refueling habits, and route planning. The power mode adopts discrete coding, including: 0-pure hydrogen fuel cell vehicle, 1-hydrogen-electric hybrid vehicle, 2-vehicle with energy recovery function; the vehicle operation mode adopts discrete coding, including: 0-normal mode, 1-economy mode, 3-sport mode; the ambient temperature, air resistance, gradient resistance, traffic conditions, and road conditions data are collected from Baidu API and converted.

3. The system as described in claim 1, characterized in that: The preprocessing of the relevant feature data specifically includes: First, vehicles with no driving records or mileage exceeding 800km are removed from the daily vehicle data. The log data is then segmented using a sliding window, with a time window of half an hour and a sliding step of 10 minutes, resulting in 143 operating condition segments for each vehicle per day. Vehicles with a missing feature rate greater than 10% in any of these segments are removed. Then, for a given vehicle's historical data, its operating condition segments are divided into 1000 classes using a clustering algorithm and features without missing values. Missing values ​​are corrected using interpolation with complete operating condition segments within the same class. The formula used is: The eigenvalues ​​are normalized to the range [0,1], where These are the original eigenvalues. This is the maximum value of the characteristic. The minimum value of this feature is used; the features of power mode, vehicle operation mode, and road type of route planning are encoded using one-hot encoding.

4. The system as described in claim 3, characterized in that: The fuel cell stack energy consumption prediction model specifically employs a recurrent neural network based on the GRU model, and the corresponding immediate benefits... The calculation method is as follows: First, construct a GRU model, train the model using the feature set selected each time, and use the feature vector of the test set corresponding to the feature set selected each time to make a prediction, and calculate the deviation between the predicted hydrogen consumption and the actual hydrogen consumption. Where m is the number of samples in the test set. This is a collection of hydrogen consumption rates over a half-hour period. These are the predicted values ​​from the GRU model.

5. The system as described in claim 4, characterized in that: Using the selected feature set X, the GRU model performs the following computation process: in, The input data at time t is taken from the features. , , These are the logic units for the update gate and the reset gate, respectively. Control the network to discard or add information from previous time series. The control network determines the extent to which information is updated; The hidden layer parameters of the network at time t-1, for The predicted value, , , , These are the weight matrices for the update gate, reset gate, state unit, and output layer, respectively. Here is the predicted hydrogen fuel consumption at time t. Let be the Sigmoid activation function and tanh be the tanh activation function. They are calculated as follows: ； A loss function is defined for the GRU model. The model parameters are updated iteratively using backpropagation with the objective of minimizing the model loss function. , , , This continues until the model converges; the loss function of the model is designed as follows: During model training, the data in the feature set is first sorted by history, and then it is traversed step by step with multiple continuous working condition time windows as the length to obtain a series of continuous working condition segments. The last working condition data of the segment is used as the predicted value, and the first few working condition data of the segment are used as the feature values.

6. The system as described in claim 1, characterized in that: The energy consumption prediction model module specifically uses the PyTorch deep learning framework for model training, verification, and testing. By preprocessing the test set data using the same methods, standardized data input is obtained. The trained model file is loaded and run to obtain the predicted values ​​of the test data. The mean squared error between the predicted and true values ​​is calculated using the loss function formula to measure the model's generalization ability.

Images