Energy storage optimal configuration method and system based on data-driven model fusion

Abstract

The invention discloses an energy storage optimal configuration method and system based on data-driven model fusion. The method comprises the following steps: extracting feature data with important influence in load data and operation and maintenance data for data fusion; constructing a load prediction model for predicting the load state of the energy storage equipment and a predictive maintenance model for predicting the health state and the remaining service life state of the energy storage equipment; training a load prediction model and a predictive maintenance model of the energy storage system through iteration; and the load state, the health state and the remaining service life state of the energy storage equipment are predicted in real time, an energy storage system configuration optimization model is constructed, and the overall configuration of the energy storage system is optimized. The method is not only used for short-term power load prediction, but also used for health state prediction and residual service life prediction of the energy storage system, through the accurately predicted energy storage equipment state, an energy storage system configuration optimization model is cooperatively constructed in a complex power grid scene, a maintenance plan is reasonably arranged, and efficient and stable operation of the energy storage system is ensured; and the reliability and the economical efficiency of power supply are improved.

Description

Technical Field

[0001] The present invention belongs to the field of energy storage optimization, and particularly relates to a method and system for optimizing the configuration of energy storage based on data-driven model fusion. Background Art

[0002] The power grid transmission and distribution system faces many challenges, such as load fluctuations, uncertainties in new energy power generation, etc. As an important part of the transmission and distribution system, the energy storage system can flexibly configure the energy supply, quickly respond to the changes in the power grid, store electrical energy during low electricity price or power surplus periods, and release electrical energy during peak periods to balance the grid load, effectively alleviating the tension in power supply and improving the reliability and stability of power supply. In addition, the energy storage system can also solve the output volatility and intermittency of wind power and photovoltaic power generation, improve the utilization rate of clean energy, and enhance the grid's acceptance capacity for renewable energy power generation grid connection.

[0003] The optimal configuration of the energy storage system is of great significance for improving the overall performance of the power grid, including ensuring the stable operation of the power grid under different load conditions by quickly responding to grid demands and providing auxiliary services; improving the operation efficiency and economic benefits of the power market by participating in power market transactions and dispatching, and providing strong support for the stable operation of the power market; providing emergency backup power for important loads in case of emergencies or extreme situations, enhancing the resilience and emergency guarantee ability of the power system, and ensuring power supply safety.

[0004] Traditional methods for optimizing the configuration of energy storage include the differential supplementary method, the smoothing analysis method, and the economic optimization method. Among them, the differential supplementary method does not consider the dynamic change process of the energy storage system capacity in actual operation, and the capacity configuration is often not accurate enough; the smoothing analysis method does not consider the quantitative analysis of the energy storage system capacity, and the capacity configuration is often affected by power disturbances, with large errors; although the economic optimization method comprehensively considers various factors, it is often difficult to balance the accuracy and complexity of its model. An overly complex model may lead to difficult solutions, while an overly simple model cannot accurately reflect the actual situation. Therefore, in practical applications, various factors need to be comprehensively considered to select a suitable technical method for optimizing the configuration of energy storage. Summary of the Invention

[0005] The object of the present invention is to provide a method and system for optimizing the configuration of energy storage based on data-driven model fusion in view of the above problems existing in the prior art.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] In the first aspect, the present invention proposes a method for optimizing the configuration of energy storage based on data-driven model fusion, including:

[0008] S1. Collect the load data and operation and maintenance data of the energy storage system, respectively extract the characteristic data that has an important impact on energy storage load prediction and equipment status prediction, and perform data fusion;

[0009] S2. Based on the fused characteristic data for energy storage load prediction, construct a load prediction model for predicting the load status of energy storage equipment; based on the fused characteristic data for equipment status prediction, construct a predictive maintenance model for predicting the health status and remaining service life status of energy storage equipment;

[0010] S3. Iteratively train the load prediction model and the predictive maintenance model of the energy storage system to obtain the trained load prediction model and predictive maintenance model;

[0011] S4. Use the trained load prediction model and predictive maintenance model to predict the load status, health status and remaining service life status of energy storage equipment in real time, construct an energy storage system configuration optimization model, and optimize the overall configuration of the energy storage system.

[0012] In the above S2, the specific steps for constructing the load prediction model include:

[0013] A. Based on the fused characteristic data for energy storage load prediction, use the following formula to calculate the retained information and new information in the load prediction model:

[0014] F t =σ(W f ·[H t-1 , X t +b f );

[0015] I t =σ(W i ·[H t-1 , X t +b i );

[0016]

[0017] In the above formula, F t is the retained information in the load prediction model at the current moment, σ is the Sigmoid function, W f is the weight matrix of the retained information, [H t-1 , X t is the concatenation of the previous moment's hidden state and the current moment's input data, b f is the bias term of the retained information, I t is the new information in the load prediction model at the current moment, W i is the weight matrix of the new information, b i is the bias term of the new information, is the candidate value vector containing new information at the current moment, tanh is the hyperbolic tangent function, and W c is the weight matrix of the candidate value vector, and b c is the bias term of the candidate value vector;

[0018] B. Combining the retained information and the new information in the load prediction model, update the input of the energy storage system load prediction model using the following formula:

[0019]

[0020] In the above formula, C t is the input of the updated energy storage system load prediction model at the current moment, and C t-1 is the input of the updated energy storage system load prediction model at the previous moment;

[0021] C. Based on the input of the updated energy storage system load prediction model, calculate the hidden state output at the current moment using the following formula:

[0022] O t =σ(W o ·[H t-1 , X t +b o );

[0023] H t =O t ·tanh(C t );

[0024] In the above formula, O t is the output information at the current moment, W o is the weight matrix of the output information, b o is the bias term of the output information, and H t is the hidden state output at the current moment.

[0025] In step S2, the specific steps for constructing the predictive maintenance model include:

[0026] a. Capture the spatial and temporal dependencies between the device status prediction feature data, and generate the spatial self-attention weight and the temporal self-attention weight:

[0027] α = softmax(S);

[0028] β = softmax(U);

[0029]

[0030] In the above formula, α is the spatial self-attention weight, S is the spatial dependence relationship between feature data, β is the temporal self-attention weight, U is the temporal dependence relationship between feature data, Vs is the input feature matrix of spatial self-attention, Ws and Rs are both learnable weight matrices for capturing spatial feature dependence relationships, Vt is the input feature matrix of temporal self-attention, Wt and Rt are both independent weight matrices for capturing temporal feature dependence relationships;

[0031] b. Further extract the spatial and temporal dynamic relationships between feature data using convolution to obtain the corresponding convolved feature matrix:

[0032] V′ I = A′ * (αV);

[0033]

[0034] In the above formula, V′ I is the convolved feature matrix in space, A′ is the adjacency matrix after self-attention weighting, V is the input feature matrix, is the convolved feature matrix in time, Mish is the activation function, W H is the weight matrix of temporal convolution, * is the convolution operation.

[0035] The S3 includes:

[0036] S31. Considering the characteristics of the energy storage system and the specific requirements of the prediction task, different samples are weighted to construct the weighted loss functions of the load prediction model and the predictive maintenance model of the energy storage system;

[0037] The weighted loss function of the load prediction model is:

[0038]

[0039] In the above formula, L δ (y1, f(x)1) is the loss function of the load prediction model, M is the total number of load prediction samples, w i is the weight of the load prediction sample i, y i，1 is the true value of the load prediction sample i, f(x i )1 is the load prediction value of the load prediction sample i, δ is a hyperparameter used to control the sensitivity of the loss function to outliers;

[0040] The weighted loss function of the predictive maintenance model is:

[0041]

[0042] In the above formula, L τ(y2, f(x)2) is the loss function of the predictive maintenance model, N is the total number of predictive maintenance samples, and w j is the weight of the predictive maintenance sample j, y j，2 is the true value of the predictive maintenance sample j, f(x j )2 is the predicted value of the predictive maintenance sample j, and τ is the hyperparameter of the quantile;

[0043] S32. Randomly select a small batch of samples and calculate the gradients of the loss functions of the load prediction model and the predictive maintenance model;

[0044] The gradient of the loss function of the load prediction model is:

[0045]

[0046] In the above formula, is the gradient of the loss function of the load prediction model with respect to the current model parameters, is the gradient of the load prediction value f(x)1 with respect to the model parameter θt;

[0047] The gradient of the loss function of the predictive maintenance model is:

[0048]

[0049] In the above formula, is the gradient of the loss function of the predictive maintenance model with respect to the current model parameters, is the gradient of the load prediction value f(x)2 with respect to the model parameter ζt;

[0050] S33. According to the gradients of the loss functions of the load prediction model and the predictive maintenance model with respect to the corresponding model parameters, update the model parameters using the following formula:

[0051]

[0052] In the above formula, θ t+1 is the updated parameter of the load prediction model, η is the learning rate, which is used to control the step size of parameter update, and ζ t+1 is the updated parameter of the predictive maintenance model;

[0053] S34. Return to step S31, repeat the iterative calculation of the loss function and the corresponding gradients to update the model parameters until the model parameters converge or reach the preset number of iterations, and the training of the load prediction model and the predictive maintenance model ends.

[0054] In the above S4, the objective function of the energy storage system configuration optimization model includes:

[0055] min(C energy +C maintenance +Creplacement +C penalty );

[0056] C energy = λ(t)P grid (t)Δt;

[0057]

[0058] C penalty = γ|D(t) - P discharge (t)|;

[0059] In the above formula, C energy is the energy cost, C maintenance is the maintenance cost, C replacem ent is the equipment replacement cost, C penalty is the penalty cost, λ(t) is the time-of-use electricity price at time t, P grid (t) is the electricity purchased from the power grid at time t, Δt is the time variation, α and β are the fixed cost coefficients of the maintenance cost and the replacement cost respectively, is the maintenance indicator function, which is 1 when the maintenance action is performed and 0 otherwise, is the replacement indicator function, which is 1 when the replacement action is performed and 0 otherwise, γ is the penalty coefficient for supply-demand imbalance, D(t) is the load demand at time t, P discharge (t) is the discharge power of the energy storage system at time t;

[0060] The constraint conditions of the energy storage system configuration optimization model include the energy storage physical limit constraint, the power grid supply-demand balance constraint, and the equipment health and maintenance constraint;

[0061] The energy storage physical limit constraint includes:

[0062] 0 ≤ SOC(t) ≤ C storage ;

[0063] 0 ≤ P charge (t) ≤ P max ;

[0064] 0 ≤ P discharge (t) ≤ P max ;

[0065]

[0066] In the above formula, SOC(t) is the state of charge of the energy storage system at time t, C storage is the energy storage capacity, which is dynamically adjusted according to the long-term load prediction result, P charge (t), P discharge (t) are the charge and discharge powers of the energy storage system at time t respectively, P maxis the maximum rated power of the energy storage device, η charge and η discharge are the charging and discharging efficiencies of the energy storage system at time t, respectively;

[0067] The power grid supply-demand balance constraint is:

[0068] P discharge (t) - P charge (t) + P grid (t) = D(t);

[0069] In the above formula, P grid (t) is the power grid power purchase at time t, and D(t) is the load demand at time t;

[0070] The equipment health and maintenance constraints include:

[0071]

[0072] In the above formula, SOH(t) is the predicted value of the health state at time t, SOH critical is the health state threshold, t maintenance is the maintenance plan time point. When the predicted result of the health state is lower than the threshold, the energy storage device is arranged to stop for maintenance. Δt delay is the delay time, RUL(t) is the predicted value of the remaining useful life at time t, and RUL min is the minimum threshold of the remaining useful life, t replace is the energy storage device replacement plan. When the predicted result of the remaining useful life is close to the end of life, the device replacement process is triggered. Δt replace is the advance ordering time.

[0073] On the second aspect, the present invention proposes an energy storage optimization configuration system based on data-driven model fusion, including a data fusion module, a model construction module, a model training module, and a configuration optimization module;

[0074] The data fusion module is used to collect the load data and operation and maintenance data of the energy storage system, and respectively extract the characteristic data that has an important impact on the energy storage load prediction and device state prediction for data fusion;

[0075] The model construction module is used to construct a load prediction model for predicting the load state of the energy storage device based on the fused energy storage load prediction characteristic data; and construct a predictive maintenance model for predicting the health state and remaining useful life state of the energy storage device based on the fused device state prediction characteristic data;

[0076] The model training module is used to iteratively train the load prediction model and the predictive maintenance model of the energy storage system to obtain the trained load prediction model and predictive maintenance model;

[0077] The configuration optimization module is used to utilize the trained load prediction model and predictive maintenance model to predict the load status, health status, and remaining service life status of the energy storage device in real time, construct an energy storage system configuration optimization model, and optimize the overall configuration of the energy storage system.

[0078] The model construction module includes a load prediction model construction unit;

[0079] The load prediction model construction unit includes an information calculation subunit, an input update subunit, and a hidden state output calculation subunit;

[0080] The information calculation subunit is used to calculate the retained information and new information in the load prediction model based on the fused energy storage load prediction feature data using the following formula:

[0081] F t =σ(W f ·[H t-1 ,X t +b f );

[0082] I t =σ(W i ·[H t-1 ,X t +b i );

[0083]

[0084] In the above formula, F t is the retained information in the load prediction model at the current moment, σ is the Sigmoid function, W f is the weight matrix of the retained information, [H t-1 ,X t is the concatenation of the hidden state at the previous moment and the input data at the current moment, b f is the bias term of the retained information, I t is the new information in the load prediction model at the current moment, W i is the weight matrix of the new information, b i is the bias term of the new information, is the candidate value vector containing new information at the current moment, tang is the hyperbolic tangent function, W c is the weight matrix of the candidate value vector, b c is the bias term of the candidate value vector;

[0085] The input update subunit is used to combine the retained information and new information in the load prediction model and update the input of the energy storage system load prediction model using the following formula:

[0086]

[0087] In the above formula, C t is the input of the energy storage system load prediction model updated at the current moment, and C t-1 is the input of the energy storage system load prediction model updated at the previous moment;

[0088] The hidden state output calculation subunit is used to calculate the hidden state output at the current moment based on the input of the updated energy storage system load prediction model by using the following formula:

[0089] O t = σ(W o · [H t-1 , X t + b o );

[0090] H t = O t · tanh(C t );

[0091] In the above formula, O t is the output information at the current moment, W o is the weight matrix of the output information, b o is the bias term of the output information, and H t is the hidden state output at the current moment.

[0092] The model construction module further includes a predictive maintenance model construction unit;

[0093] The predictive maintenance model construction unit includes a self-attention weight generation subunit and a feature matrix convolution subunit;

[0094] The self-attention weight generation subunit is used to capture the spatial and temporal dependence relationships between the device state prediction feature data and generate spatial self-attention weights and temporal self-attention weights:

[0095] α = softmax(S);

[0096] β = softmax(U);

[0097]

[0098] In the above formula, α is the spatial self-attention weight, S is the spatial dependence relationship between the feature data, β is the temporal self-attention weight, U is the temporal dependence relationship between the feature data, Vs is the input feature matrix of the spatial self-attention, Ws and Rs are both learnable weight matrices for capturing the spatial feature dependence relationship, Vt is the input feature matrix of the temporal self-attention, Both Wt and Rt are independent weight matrices that capture the time - feature dependence relationship;

[0099] The feature matrix convolution sub - unit is used to further extract the spatial and temporal dynamic relationships between feature data by convolution, obtaining the corresponding convolved feature matrix:

[0100] V′I = A′ * (αV);

[0101]

[0102] In the above formula, V′I is the feature matrix after spatial convolution, A′ is the adjacency matrix after self - attention weighting, V is the input feature matrix, is the feature matrix after temporal convolution, Mish is the activation function, WH is the weight matrix of temporal convolution, and * is the convolution operation.

[0103] The model training module includes a loss function construction unit, a gradient calculation unit, a model parameter update unit, and an iterative training unit;

[0104] The loss function construction unit is used to consider the characteristics of the energy storage system and the specific requirements of the prediction task, weight different samples, and construct the weighted loss functions of the load prediction model and the predictive maintenance model of the energy storage system;

[0105] The weighted loss function of the load prediction model is:

[0106]

[0107] In the above formula, L δ (y1, f(x)1) is the loss function of the load prediction model, M is the total number of load prediction samples, w i is the weight of the load prediction sample i, y i，1 is the true value of the load prediction sample i, f(x i )1 is the load prediction value of the load prediction sample i, and δ is a hyperparameter used to control the sensitivity of the loss function to outliers;

[0108] The weighted loss function of the predictive maintenance model is:

[0109]

[0110] In the above formula, L τ (y2, f(x)2) is the loss function of the predictive maintenance model, N is the total number of predictive maintenance samples, w j is the weight of the predictive maintenance sample j, y j，2 is the true value of the predictive maintenance sample j, f(x j)2 is the predicted value of the predictive maintenance sample j, and τ is the hyperparameter of the quantile;

[0111] The gradient calculation unit is used to randomly select a small batch of samples and calculate the gradients of the loss functions of the load prediction model and the predictive maintenance model;

[0112] The gradient of the loss function of the load prediction model is:

[0113]

[0114] In the above formula, is the gradient of the loss function of the load prediction model with respect to the current model parameters, is the gradient of the load prediction value f(x)1 with respect to the model parameter θt;

[0115] The gradient of the loss function of the predictive maintenance model is:

[0116]

[0117] In the above formula, is the gradient of the loss function of the predictive maintenance model with respect to the current model parameters, is the gradient of the load prediction value f(x)2 with respect to the model parameter ζt;

[0118] The model parameter update unit is used to update the model parameters according to the gradients of the loss functions of the load prediction model and the predictive maintenance model with respect to the corresponding model parameters, using the following formula:

[0119]

[0120] In the above formula, θ t+1 is the updated load prediction model parameter, η is the learning rate, used to control the step size of parameter update, ζ t+1 is the updated predictive maintenance model parameter

[0121] The iterative training unit is used to return to the loss function construction unit, repeatedly calculate the loss function and the corresponding gradients to update the model parameters until the model parameters converge or reach the preset number of iterations, and the training of the load prediction model and the predictive maintenance model ends.

[0122] In the configuration optimization module, the objective function of the energy storage system configuration optimization model includes:

[0123] min(C energy +C maintenance +C replacement +C penalty );

[0124] C energy =λ(t)Pgrid (t)Δt;

[0125]

[0126] C penalty =γ|D(t) - P discharge (t)|;

[0127] In the above formula, C energy is the energy cost, C maintenance is the maintenance cost, C replacement is the equipment replacement cost, C penalty is the penalty cost, λ(t) is the time-of-use electricity price at time t, P grid (t) is the electricity purchased from the grid at time t, Δt is the time variation, α and β are the fixed cost coefficients of the maintenance cost and the replacement cost respectively, is the maintenance indication function, which is 1 when the maintenance action is carried out and 0 otherwise, is the replacement indication function, which is 1 when the replacement action is carried out and 0 otherwise, γ is the penalty coefficient for supply-demand imbalance, D(t) is the load demand at time t, P discharge (t) is the discharge power of the energy storage system at time t;

[0128] The constraint conditions of the energy storage system configuration optimization model include the energy storage physical limit constraint, the grid supply-demand balance constraint, and the equipment health and maintenance constraint;

[0129] The energy storage physical limit constraint includes:

[0130] 0 ≤ SOC(t) ≤ C storage ;

[0131] 0 ≤ P charge (t) ≤ P max ;

[0132] 0 ≤ P discharge (t) ≤ P max ;

[0133]

[0134] In the above formula, SOC(t) is the state of charge of the energy storage system at time t, C storage is the energy storage capacity, and the energy storage capacity is dynamically adjusted according to the long-term load prediction result, P charge (t), P discharge (t) are the charge and discharge powers of the energy storage system at time t respectively, P max is the maximum rated power of the energy storage equipment, η charge , η discharge are the charge and discharge efficiencies of the energy storage system at time t respectively;

[0135] The power grid supply-demand balance constraint is as follows:

[0136] P discharge (t) - P charge (t) + P grid (t) = D(t);

[0137] In the above formula, P grid (t) is the power grid power purchase at time t, and D(t) is the load demand at time t;

[0138] The equipment health and maintenance constraints include:

[0139]

[0140] In the above formula, SOH(t) is the predicted value of the health state at time t, SOH critical is the health state threshold, t maintenance is the maintenance plan time point. When the predicted result of the health state is lower than the threshold, the energy storage device is arranged for shutdown maintenance, Δt delay is the delay time, RUL(t) is the predicted value of the remaining useful life at time t, RUL min is the minimum threshold of the remaining useful life, t replace is the energy storage device replacement plan. When the predicted result of the remaining useful life is close to the end of the life, the device replacement process is triggered, Δt replace is the advance order time.

[0141] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0142] The present invention provides a method and system for optimizing the configuration of energy storage based on data-driven model fusion. The method first collects the load data and operation and maintenance data of the energy storage system, extracts the characteristic data that has an important impact on energy storage load prediction and equipment status prediction respectively for data fusion; then, based on the fused characteristic data of energy storage load prediction, a load prediction model for predicting the load status of energy storage equipment is constructed; based on the fused characteristic data of equipment status prediction, a predictive maintenance model for predicting the health status and remaining service life status of energy storage equipment is constructed; the load prediction model and the predictive maintenance model of the energy storage system are iteratively trained to obtain the trained load prediction model and predictive maintenance model; finally, the trained load prediction model and predictive maintenance model are used to predict the load status, health status and remaining service life status of energy storage equipment in real time, construct an optimization model for the overall configuration of the energy storage system, and optimize the overall configuration of the energy storage system. On the one hand, the method is aimed at the energy storage system in the power transmission and distribution system, involves multi-dimensional information such as load data and operation and maintenance data, constructs a load prediction model to focus on the external demand of the energy storage system, optimizes the charge and discharge plan, reduces the energy cost, constructs a predictive maintenance model to focus on the internal state of the energy storage equipment, reduces the emergency repair cost, extends the equipment life, has diversified data sources, wide application scenarios of the model, and more accurately predicts the state of the energy storage equipment; on the other hand, the method is not only used for short-term power load prediction, but also for the health status prediction and remaining service life prediction of the energy storage system. Through the accurately predicted state of the energy storage equipment, an optimization model for the configuration of the energy storage system is collaboratively constructed in a complex power grid scenario, the maintenance plan is reasonably arranged, the efficient and stable operation of the energy storage system is ensured, and the reliability and economy of power supply are improved. Brief Description of the Drawings

[0143] Figure 1 It is the overall flowchart of the method described in the present invention.

[0144] Figure 2 It is the structure diagram of the system described in the present invention. Detailed Embodiments

[0145] The present invention will be further described in detail below in combination with the detailed embodiments and the drawings.

[0146] The present invention proposes an energy storage optimization configuration method and system based on data-driven model fusion. Load data and operation and maintenance data are extracted from the power energy storage and management system, and key features are extracted after preprocessing. Data with different time granularities are aggregated, and the Kalman filter is used to fuse the data, improving the accuracy and reliability of the data. An LSTM network is used to construct a load prediction model, and a predictive maintenance model is constructed by combining self-attention and spatio-temporal graph convolutional networks. Considering the characteristics of the energy storage system and the specific requirements of the prediction task, a more accurate weighted loss function is constructed, the random gradient descent algorithm is used to optimize the model parameters, and dropout regularization is used to prevent overfitting. Appropriate evaluation indicators are selected to evaluate the effects of the load prediction and predictive maintenance models to verify the practicability and reliability of the models. Finally, an energy storage system configuration optimization model is constructed to optimize the overall configuration of the energy storage system.

[0147] Embodiment 1:

[0148] As Figure 1 shown, an energy storage optimization configuration method based on data-driven model fusion is carried out in sequence according to the following steps:

[0149] 1. Collect the load data and operation and maintenance data of the energy storage system for preprocessing, and extract the characteristic data that has an important impact on the energy storage load prediction and equipment status prediction respectively;

[0150] Comprehensively extract the load data from the energy storage management system of the power system, including key indicators such as the charge and discharge power and the change of state of charge (SOC) during the operation of the energy storage device for load prediction; extract the operation and maintenance data from the monitoring system or sensors of the energy storage device, including key indicators such as equipment operation parameters and health status, which can reflect the operation status and performance of the energy storage device in real time and are used for the predictive maintenance of the energy storage device;

[0151] Perform data cleaning on the collected data, including identifying and processing outliers, duplicates, and missing values to ensure the quality and accuracy of the data: for outliers, use the 3σ rule algorithm to identify, and the formula is |x - μ| > 3σ, that is, when the difference between the data point x and the mean μ exceeds 3 times the standard deviation σ, this data point is regarded as an outlier, and the methods for processing outliers include deletion, replacement, and retention but marking; for duplicates, identify by checking the unique identifier or all fields of the data and delete the duplicate records; for missing values, use interpolation method, mean / median filling method, and forward / backward value filling method to fill;

[0152] Perform data standardization on the cleaned data to convert data from different sources and different types to the same scale for comparison and analysis, and use the Z-score standardization formula to convert each data point:

[0153]

[0154] In the above formula, z is the standardized data, x is the original data point, μ is the mean, and σ is the standard deviation. Among many standardization methods, Z-score standardization is a commonly used and effective method. For each feature data, first calculate its mean and standard deviation on the entire dataset. The mean is the sum of all data points divided by the number of data points, which reflects the central position of the data. The standard deviation is a measure of the degree to which data points deviate from the mean, reflecting the dispersion of the data.

[0155] Extract feature data from the original data that has an important impact on energy storage load prediction and equipment status prediction, including the daily change rate of the load, the temperature change trend of the equipment, the charge and discharge times of the energy storage, the change relationship of the state of charge (SOC), etc. Use methods such as correlation coefficient analysis, recursive feature elimination (RFE), and feature importance evaluation based on tree models to select the most critical feature data for modeling, improving the efficiency and prediction accuracy of the model.

[0156] 2. Perform data fusion on the feature data that has an important impact on energy storage load prediction and equipment status prediction;

[0157] First, integrate feature data with different time granularities onto the same time scale through data aggregation;

[0158] The time granularities of the extracted feature data are different. Equipment operation and maintenance data may be recorded in minutes, while load data may be statistically calculated in hours. For high-frequency data, it can be aggregated into low-frequency data at the hourly level by calculating the average value or total within each hour. The formula for data aggregation can be expressed as:

[0159] Average value aggregation:

[0160] Total value aggregation:

[0161] In the above formula, A hour is the aggregated data at the hourly level, is the original feature data at the i-th minute; Organize and match the aggregated data with the time stamps of the low-frequency data to ensure that all data is on the same time scale;

[0162] Then, use the Kalman filter to fuse feature data from different sources and of different natures on the same time scale to form a comprehensive dataset to estimate the state of the energy storage device and achieve accurate modeling of the feature data.

[0163] In load forecasting, the Kalman filter fuses the measured values such as the current and voltage of the energy storage device, as well as data such as the output of the dynamic model of the energy storage device, to estimate the load state of the energy storage device; in predictive maintenance, the Kalman filter fuses information such as the operating data and historical maintenance records of the energy storage device to estimate the state of health SOH and remaining useful life state RUL of the energy storage device; the specific steps of Kalman filter fusion include:

[0164] The Kalman filter is not only used for data fusion, but more importantly, it can synthesize various factors, including state variables such as voltage, current, temperature, and charge-discharge times, and dynamically adjust the state estimation of the energy storage device in real time according to new measurement data, accurately estimate the overall state of the energy storage device, and improve the accuracy and real-time performance of the prediction;

[0165] Initialize the data state variables and error covariance matrix:

[0166]

[0167] P(0|0) = P0;

[0168] In the above formula, is the state estimation at the initial moment of the data, x0 is the initial value of the state, P(0|0) is the initial error covariance matrix, and P0 is a large diagonal matrix;

[0169] Perform prediction update on the state variables and error covariance matrix:

[0170]

[0171]

[0172] In the above formula, is the prediction based on the state at the previous moment, A is the state transition matrix, B is the control input matrix, u(k - 1) is the control input at the previous moment, P(k|k - 1) is the prediction based on the error covariance at the previous moment, is the system noise covariance matrix;

[0173] Combine the measurement value to correct the prediction update of the state variables and error covariance matrix to obtain a more accurate state estimation:

[0174]

[0175] P(k|k) = (I - K(k)C)P(k|k - 1);

[0176] K(k) = P(k|k - 1)C T (CP(k|k - 1)C T +R) -1 ;

[0177] In the above formula, is the corrected state estimate, K(k) is the Kalman gain, y(k) is the measured value at the current moment, C is the observation matrix, P(k|k) is the corrected error covariance matrix, I is the identity matrix, and R is the measurement noise covariance matrix.

[0178] 3. Based on the fused energy storage load prediction feature data, construct a load prediction model for predicting the load state of the energy storage device; based on the fused device state prediction feature data, construct a predictive maintenance model for predicting the health state and remaining service life state of the energy storage device;

[0179] When constructing the load prediction model, set a series of different numbers of hidden layer units (such as 32, 64, 128, 256, etc.) and time steps (such as 1 hour, 2 hours, 4 hours, 8 hours, etc.). Among them, the specific time step needs to be determined according to the sampling frequency of the data and the load change characteristics; for each set number of hidden layer units, loop through each set time step. Under the combination of each number of hidden layer units and time step, construct a load prediction model, and train the model according to the iterative training method described below to ensure the consistency and effectiveness of the training process. After each iteration, use the validation set to evaluate the model performance, record the key indicators of the model, analyze the change trend of the model performance under different combinations of the number of hidden layer units and time steps, and determine the optimal parameter combination;

[0180] Based on the LSTM network, through mechanisms such as cell state, forget gate, input gate, output gate, etc., construct a load prediction model. This model can efficiently process load data and accurately predict the load state of the energy storage device by capturing long-term dependencies in the time series;

[0181] The LSTM network is a special recurrent neural network RNN that can effectively capture long-term dependencies in the time series. Among them, the cell state runs through the entire network, used to record the long-term stored information in the network and is updated at each time step; the forget gate determines the degree of information retention in the cell state at the previous moment; the input gate controls the update of the cell state by the input information at the current moment; the output gate controls the output of the hidden state at the current moment. The specific steps for constructing its load prediction model include:

[0182] The cell state in the LSTM receives the feature data fused by Kalman filtering, including voltage and current data extracted from the energy storage monitoring system, charge and discharge power, state of charge SOC, etc. output by the energy storage dynamic model, as the input of the load prediction model;

[0183] The forget gate processes the input through the Sigmoid activation function to determine which information in the load prediction model should be forgotten, such as historical load data, and which information should be retained. The calculation formula for the retained information is:

[0184] F t =σ(W f ·[H t-1 ,X t +b f );

[0185] In the above formula, F t is the retained information in the load prediction model at the current time, σ is the Sigmoid function, W f is the weight matrix of the retained information, [H t-1 ,X t is the concatenation of the previous hidden state and the current input data, and b f is the bias term of the retained information;

[0186] The input gate determines which new information in the input of the load prediction model should be accepted and stored through the Sigmoid activation function, and generates a new candidate value vector through the tanh function. This vector contains new information that may be updated to the cell state. The calculation formula for the new information is:

[0187] I t =σ(W i ·[H t-1 ,X t +b i );

[0188]

[0189] In the above formula, I t is the new information in the load prediction model at the current time, W i is the weight matrix of the new information, b i is the bias term of the new information, is the candidate value vector containing new information at the current time, tanh is the hyperbolic tangent function, W c is the weight matrix of the candidate value vector, and b c is the bias term of the candidate value vector;

[0190] Combining the retained information of the load prediction model in the forget gate and the new information in the input gate, the cell state of the LSTM is updated, that is, the input of the load prediction model is updated. The old cell state is screened by the forget gate to discard unimportant information. At the same time, the new candidate value vector is weighted by the input gate, and new information, including the current current, voltage measurement values, and updated SOC estimation values, is integrated into the cell state. The calculation formula for the update is:

[0191]

[0192] In the above formula, C t is the input of the energy storage system load prediction model updated at the current moment, and C t-1 is the input of the energy storage system load prediction model updated at the previous moment;

[0193] The output gate processes the initial input through the Sigmoid activation function, determines which information should be output to the next layer or the external system, obtains the output information at the current moment, and processes the input updated by the load prediction model through the tanh function to obtain a value vector containing the information at the current moment. Multiply the value vector by the output information at the current moment to obtain the final hidden state output, which reflects the fusion between the state of the energy storage device and the load data. The calculation formula for the hidden state output at the current moment is:

[0194] O t = σ(W o · [H t-1 , X t + b o );

[0195] H t = O t · tanh(C t );

[0196] In the above formula, O t is the output information at the current moment, W o is the weight matrix of the output information, b o is the bias term of the output information, and H t is the hidden state output at the current moment;

[0197] Among them, the number of hidden layer units of the LSTM network can be initially selected within a range according to experience or preliminary experiments, the complexity of the energy storage system load data, and the difficulty of the prediction task: for more complex load data, a larger number of units can be selected, such as starting with 100 - 200 units;

[0198] The time step of the LSTM network can be selected according to the periodicity and change trend of the energy storage system load data. If the load data has an obvious daily periodicity, 24 or 48 time steps can be selected as the input;

[0199] Experiments are conducted with different numbers of hidden layer units and time steps, and the performance of the model on the validation set is observed. The number of hidden layer units and time steps that minimize the prediction error are selected as the optimal parameters. By adjusting the number of hidden layer units, the model can adapt to the load characteristics of different energy storage systems. Increasing the number of hidden layer units can enhance the learning ability of the model, enabling it to better capture complex patterns in the load data. By adjusting the time step, the model can adapt to prediction requirements at different time scales, such as short-term, medium-term, and long-term predictions. Moreover, by utilizing more historical information, the model can predict future load changes more stably and reduce the volatility of predictions.

[0200] Predictive maintenance utilizes equipment operation and maintenance feature data and, through a model based on the self-attention mechanism and spatio-temporal graph convolutional network, realizes predictive maintenance of the faults, health status, and remaining useful life of energy storage equipment.

[0201] For the fault detection of energy storage equipment, first, the real-time operation data and historical operation and maintenance data of the energy storage equipment are input into the predictive maintenance model. The spatio-temporal graph convolutional network is used to extract spatial and temporal features from the input data, capturing the complex relationships between different components of the energy storage equipment and between different time points. Then, the self-attention mechanism is used to further emphasize the features that have an important impact on the predicted faults. The self-attention mechanism can automatically learn the correlations between features and highlight the features highly relevant to fault prediction. The fully connected layer is used to map the fused features to the fault probability, outputting the probability of the equipment failing in a future period.

[0202] For the prediction of the state of health (SOH) and remaining useful life (RUL) of energy storage system equipment, similar to fault prediction, first, the real-time and historical operation and maintenance data of the equipment are input, and the spatio-temporal graph convolutional network is used for feature extraction. Then, based on the extracted features, the relationships between the state of health of the equipment learned by the model and the features, as well as the relationship between the equipment degradation law and the remaining useful life, are utilized to output the current state of health and remaining useful life of the equipment.

[0203] Predictive maintenance is a method for predicting equipment faults, health status, and remaining useful life based on equipment operation and maintenance data. Through a model based on the self-attention mechanism and spatio-temporal graph convolutional network, by capturing the spatial and temporal dependencies between the temperature, current, voltage, charge and discharge times, and other relevant features of the energy storage system, predictive maintenance of the energy storage system is realized. The specific steps for constructing the predictive maintenance model are as follows:

[0204] Perform a correlation analysis on the fused device status prediction feature data. Based on the results of the correlation analysis, select the features most valuable for the prediction task as the input feature matrix V. Capture the spatial and temporal dependence relationships between the input feature matrices through the self-attention mechanism to generate spatial self-attention weights and temporal self-attention weights:

[0205] α = softmax(S);

[0206] β = softmax(U);

[0207]

[0208] In the above formula, α is the spatial self-attention weight, S is the spatial dependence relationship between the feature data, β is the temporal self-attention weight, U is the temporal dependence relationship between the feature data, Vs is the input feature matrix of the spatial self-attention, Ws and Rs are both learnable weight matrices for capturing the spatial feature dependence relationship, Vt is the input feature matrix of the temporal self-attention, Wt and Rt are both independent weight matrices for capturing the temporal feature dependence relationship;

[0209] To capture the spatial and temporal dependence relationships between the components inside the energy storage device, design a weight matrix with a specific structure, is the query matrix, which is used to map the input features to the query spatial and temporal dimensions and calculate the spatial and temporal attention weights between the components. It is in the form of a learnable linear transformation matrix with a dimension of d in ×d hidden , d in is the dimension of the input feature matrix, d hidden is the dimension of the hidden layer. Ws and Wt are the key matrices, corresponding to the key mapping matrices in the spatial and temporal attention, which are combined with the query matrix to calculate the attention scores. Rs and Rt are the key matrices, generating the feature representations after the spatial and temporal attention;

[0210] Different components of the energy storage device have different state and performance characteristics, which are crucial for predicting the overall health status and remaining service life of the device. According to the physical connection relationships and mutual influences between different components of the energy storage device, such as battery cells, inverters, and thermal management systems, the way to adjust the spatial self-attention weight matrix is: adjust the input feature matrix V and the learnable weight matrices W and R, which can more accurately reflect the spatial dependence relationship between the components inside the energy storage device; by carefully selecting and adjusting the input feature matrix V, it can ensure that the model can make full use of the key information between different components of the energy storage device. The learnable weight matrices W and R determine the relative importance between different features in the self-attention mechanism. By adjusting these weight matrices, the model can more accurately capture the spatial dependencies among the components inside the energy storage device;

[0211] According to the historical operation and maintenance data of the energy storage device, including the time series characteristics such as the number of charge and discharge cycles, temperature change trend, SOC change, etc., the way to adjust the time self-attention weight matrix is as follows: adjusting the size of the time window can more accurately capture the change trend of the energy storage device's performance over time, and adjusting the learnable weight matrix can adapt to the changes in the state characteristics of the energy storage device in different time periods; the performance state of the energy storage device changes over time. Selecting an appropriate time window can ensure that the model can capture these changes, and by adjusting the size of the time window, the model can pay more attention to the time periods that have a greater impact on the prediction results; the state characteristics of the energy storage device may be different in different time periods, so the learnable weight matrix also needs to be adjusted accordingly. By making the weight matrix adapt to the feature changes in different time periods, the prediction ability of the model in the time dimension can be improved;

[0212] Use spatial and temporal convolutions to further extract the spatial and temporal dynamic relationships between feature data, and obtain the corresponding convolved feature matrix:

[0213] V′ I = A′ * (αV);

[0214]

[0215] In the above formula, V′ I is the feature matrix after spatial convolution, A′ is the adjacency matrix after self-attention weighting. Considering the dynamic changes in the actual operation of the energy storage device, such as component failures, performance degradation, etc., according to the spatial layout and mutual influence among the components of the energy storage device, the adjacency matrix A′ after self-attention weighting is updated in real time. V is the input feature matrix, is the feature matrix after temporal convolution, Mish is the activation function, which is used to increase the non-linear expression ability of the model, W H is the weight matrix of temporal convolution, and * is the convolution operation;

[0216] There are certain physical connection relationships among the components of the energy storage device. These relationships are crucial for understanding the overall structure and working principle of the device. Therefore, by adjusting the adjacency matrix A′ through the physical connection relationships, the graph convolutional network can better capture the spatial features inside the energy storage device. And the energy storage device may be affected by various factors during actual operation, resulting in changes in the connection relationships and mutual influences among the components. By updating the adjacency matrix A′ in real time, it can ensure that the graph convolutional network can accurately reflect these changes and improve the accuracy of model prediction;

[0217] According to the time series characteristics of the operation and maintenance data of the energy storage device, the weight matrix W of the temporal convolution can be adjusted by adjusting the size and stride of the convolution kernel H , so as to more accurately capture the change characteristics of the energy storage device performance over time.

[0218] 4. Iteratively train the load prediction model and the predictive maintenance model of the energy storage system to obtain the trained load prediction model and predictive maintenance model, and ensure the accuracy of the model prediction;

[0219] Considering the characteristics of the energy storage system and the specific requirements of the prediction task, different samples are weighted to construct a weighted loss function for the load prediction model and the predictive maintenance model of the energy storage system, measure the difference between the model prediction value and the true value, and improve the prediction performance of the model;

[0220] The loss function of the load prediction model is a weighted Huber loss function:

[0221]

[0222] In the above formula, L δ (y1, f(x)1) is the loss function of the load prediction model, M is the total number of load prediction samples, w i is the weight of the load prediction sample i. The weighting rule for sample i is: for the outlier sample i that significantly deviates from the normal range, a lower weight is set. For the sample i with special significance or key time points (such as peak load periods, moments of sudden changes in the state of the energy storage system, etc.), a higher weight is set, and according to the time range requirements of the prediction task, a higher weight is set for the recent sample i and a lower weight is set for the far - term sample i. y i，1 is the true value of the load prediction sample i, f(x i )1 is the load prediction value of the load prediction sample i, and δ is a hyperparameter used to control the sensitivity of the loss function to outliers. When the absolute value of the difference between the prediction value and the true value is less than or equal to δ, the loss function behaves as the mean square error. When the absolute value of the difference between the prediction value and the true value is greater than δ, the loss function behaves as a linear function;

[0223] For the load prediction of the energy storage system, using the Huber loss function can improve the robustness of the model to outliers. Especially when there are a large number of noises or outliers in the load data, the Huber loss function can more effectively guide the model training;

[0224] The loss function of the predictive maintenance model is a weighted Quantile regression loss function:

[0225]

[0226] In the above formula, L τ (y2, f(x)2) is the loss function of the predictive maintenance model, N is the total number of predictive maintenance samples, w j is the weight of predictive maintenance sample j. The weighting rule for sample j is as follows: for sample j with historical failure records or close to the failure threshold, a higher weight is set; for sample j in the normal operation state of the device, a lower weight is set; and according to the aging degree or usage time of the device, a higher weight is set for device sample j with a higher aging degree. y j，2 is the true value of predictive maintenance sample j, f(x j )2 is the predicted value of predictive maintenance sample j, and τ is the hyperparameter of the quantile, with a value range of [0, 1];

[0227] For the predictive maintenance of the energy storage system, the Quantile regression loss function is used to capture the distribution characteristics of the data. By predicting the failure time or state at different quantiles, the health status of the device can be evaluated more comprehensively, providing strong support for maintenance decisions and being applicable to evaluating the uncertainty of the prediction results;

[0228] For randomly selected samples or mini-batch samples, calculate the gradients of the loss functions of the load prediction model and the predictive maintenance model;

[0229] The gradient of the load prediction model loss function is:

[0230]

[0231] In the above formula, is the gradient of the load prediction model loss function with respect to the current model parameters, is the gradient of the load prediction value f(x)1 with respect to the model parameter θt;

[0232] The gradient of the predictive maintenance model loss function is:

[0233]

[0234] In the above formula, is the gradient of the predictive maintenance model loss function with respect to the current model parameters, is the gradient of the load prediction value f(x)2 with respect to the model parameter ζt;

[0235] Select the Stochastic Gradient Descent (SGD) algorithm to optimize and train the energy storage system model. According to the gradients of the loss functions of the load prediction model and the predictive maintenance model with respect to the corresponding model parameters, update the model parameters using the following formula:

[0236]

[0237] In the above formula, θt+1 is the updated parameter of the load prediction model, η is the learning rate, which is used to control the step size of parameter update. The choice of the learning rate has a significant impact on the performance of SGD. If it is too large, it may cause the algorithm to diverge; if it is too small, it may lead to an overly slow convergence speed. is the gradient of the loss function of the load prediction model with respect to the current model parameters, ζ t+1 is the updated parameter of the predictive maintenance model. is the gradient of the loss function of the predictive maintenance model with respect to the current model parameters;

[0238] To minimize the loss function, in each iteration, a new sample or a mini-batch of samples is randomly selected, and the iteration is repeated to calculate the loss function and the corresponding gradient to update the model parameters until the model parameters converge or reach the preset number of iterations, and the training of the load prediction model and the predictive maintenance model ends.

[0239] During the training process of the model, the dropout regularization method is adopted to prevent the model from overfitting by randomly discarding a part of neurons and their connections.

[0240] In each training iteration, for each neuron in the network, a part of neurons is randomly selected to be discarded with the probability of the Dropout ratio, that is, its output is set to 0. These discarded neurons do not participate in the forward propagation and backward propagation in the current training iteration. Among them, the Dropout ratio is a hyperparameter, which represents the proportion of randomly discarded neurons during the training process, usually set between 0.3 and 0.5, and can also be adjusted according to the complexity of the model and the characteristics of the data.

[0241] The trained predictive maintenance model will process the input device operation and maintenance data and output feature values, which reflect the current health status of the energy storage device. However, these output feature values are high-dimensional feature data extracted after spatial convolution and temporal convolution, and need to be further mapped to the predicted value of the health state SOH through a fully connected layer or other regression methods. The specific operation method is as follows: First, according to the specific requirements of the prediction task, adjust the number of layers, the number of neurons, and the weight matrix of the fully connected layer, and improve the expression ability of the model by increasing the number of layers or neurons; then, according to the specifications provided by the device manufacturer, historical operation data, and industry standards, set a series of health state thresholds, map the output features of the model to the preset health state thresholds, and determine the corresponding health state SOH of the device.

[0242] In addition to the current health status of the equipment, the predictive maintenance model outputs eigenvalue related to the degradation trend of the equipment. These eigenvalues are combined with information such as historical maintenance records, the average life of the equipment, and the mean time between failures to construct a degradation model of the equipment. According to the performance status of the current energy storage equipment, the time required for the equipment to reach the performance limit, that is, the time required for the equipment to fail, is predicted, which is the remaining useful life (RUL) of the energy storage equipment.

[0243] 5. Select appropriate evaluation indicators according to specific tasks to evaluate the effects of load forecasting and predictive maintenance;

[0244] For the load forecasting model, common evaluation indicators include mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), etc.;

[0245] The mean square error (MSE) is the average of the squares of the differences between the predicted values and the true values, reflecting the overall magnitude of the prediction error. The smaller the value of MSE, the higher the prediction accuracy of the model:

[0246]

[0247] In the above formula, n is the number of samples, y i is the true value, is the predicted value;

[0248] The mean absolute error (MAE) is the average of the absolute values of the differences between the predicted values and the true values, which reflects the average magnitude of the prediction error:

[0249]

[0250] The root mean square error (RMSE) is the square root of MSE, having the same dimension as the original data, so it is more convenient for understanding and comparison. The smaller the value of RMSE, the higher the prediction accuracy of the model:

[0251]

[0252] For the predictive maintenance model, common evaluation indicators include accuracy, precision, recall, F1-score, etc.;

[0253] Accuracy measures the proportion of the number of samples correctly predicted by the model in the total number of samples, reflecting the overall prediction performance of the model:

[0254]

[0255] In the above formula, TP is the true positive, that is, the number of samples that the model correctly predicts as the positive class; TN is the true negative, that is, the number of samples that the model correctly predicts as the negative class; FP is the false positive, that is, the number of samples that the model incorrectly predicts as the positive class; FN is the false negative, that is, the number of samples that the model incorrectly predicts as the negative class;

[0256] Precision measures the proportion of samples predicted as the positive class that are actually the positive class, reflecting the accuracy of the model's prediction of the positive class:

[0257]

[0258] Recall measures the proportion of samples that are actually the positive class and are correctly predicted as the positive class, reflecting the model's ability to capture positive class samples:

[0259]

[0260] The F1 score is the harmonic mean of precision and recall, used to balance the importance of precision and recall. Especially when there is a conflict between the two, the larger the value of the F1 score, the better the performance of the model:

[0261]

[0262] 6. Use the trained load prediction model and predictive maintenance model to predict the load status, health status, and remaining service life status of the energy storage device in real time, construct an energy storage system configuration optimization model, and optimize the overall configuration of the energy storage system;

[0263] Use the trained model to predict the load status, health status, and remaining service life of the energy storage device in real time, and optimize the overall configuration of the energy storage system by comprehensively considering these data. This includes reasonably arranging the charging and discharging plans of the energy storage system according to the load prediction results to meet the needs of the power system, adapt to the rapid changes in the load, and reduce the operating cost. Make a maintenance plan in advance according to the health status prediction results to prevent and handle potential faults. Strengthen the equipment monitoring and maintenance efforts during peak load periods or when the equipment status is poor. Reasonably arrange equipment updates or expansions according to the remaining service life prediction results to ensure the long-term efficient and stable operation of the energy storage system, realize refined management and optimization of the configuration, and improve the reliability and economy of power supply;

[0264] The specific process of optimizing the overall configuration of the energy storage system is as follows:

[0265] Define the optimization goal and decision variables of the energy storage system. Among them, the optimization goal is to balance economy and reliability by minimizing the total cost of the energy storage system; the decision variables are the core control parameters of the optimization goal, which directly affect the total cost of the energy storage system, including: charge and discharge power sequence P charge(t) and P discharge (t), determine when to charge / discharge to match the load demand; the expansion amount of energy storage capacity ΔC storage , according to the long-term load prediction results, dynamically adjust the energy storage capacity to adapt to the future load demand growth; the maintenance plan time point t maintenance , when the predicted health state result is lower than the threshold, arrange for the energy storage device to stop for maintenance; the energy storage device replacement plan t replace , when the predicted remaining service life result is close to the end of life, trigger the device replacement process to ensure timely intervention before the critical threshold;

[0266] The objective function of the energy storage system configuration optimization model includes:

[0267] min(C energy +C maintenance +C replacement +C penalty );

[0268] C energy =λ(t)P grid (t)Δt;

[0269]

[0270] C penalty =γ|D(t)-P discharge (t)|;

[0271] In the above formula, C energt is the energy cost. According to the time-of-use electricity price, charge during the low electricity price period and discharge during the high electricity price period, and use the electricity price difference to reduce the energy cost; C maintenane is the maintenance cost. According to the predicted health state result, trigger preventive maintenance before the device performance deteriorates to avoid high maintenance costs caused by sudden failures; C replacement is the device replacement cost. Based on the predicted remaining service life result, plan the device replacement time in advance to avoid power outages caused by sudden device failures; C penalty is the penalty cost, which is the grid fine or user compensation generated due to unreasonable configuration of the energy storage device, resulting in insufficient power supply or device overload; λ(t) is the time-of-use electricity price at time t, which directly affects the charge and discharge strategy, P grid (t) is the grid power purchase at time t, Δt is the time change amount, α and β are the fixed cost coefficients of the maintenance cost and replacement cost respectively, adjusted according to the device type, is the maintenance indicator function, which is 1 when a maintenance action is performed, otherwise 0, is the replacement indicator function, which is 1 when a replacement action is performed, otherwise 0, γ is the supply-demand imbalance penalty coefficient, used to forcefully meet the load demand, D(t) is the load demand at time t, P discharge(t) is the discharge power of the energy storage system at time t;

[0272] The constraint conditions of the energy storage system configuration optimization model are divided into energy storage physical limit constraints, power grid supply-demand balance constraints, and equipment health and maintenance constraints;

[0273] The energy storage physical limit constraints include:

[0274] To ensure that the state of charge (SOC) of the energy storage system is always within a safe range and avoid overcharging or over-discharging, the capacity constraint is:

[0275] 0 ≤ SOC(t) ≤ C storage ;

[0276] Limited by the hardware performance, the charge-discharge power limit constraint that the charge-discharge power shall not exceed the rated maximum value of the equipment is:

[0277] 0 ≤ P charge (t) ≤ P max ;

[0278] 0 ≤ P discharge (t) ≤ P max ;

[0279] Considering the charge-discharge efficiency, the energy conservation constraint to ensure the accuracy of the dynamic change of SOC:

[0280]

[0281] In the above formula, SOC(t) is the state of charge of the energy storage system at time t, and C storage is the energy storage capacity, and P charge (t), P discharge (t) are the charge and discharge powers of the energy storage system at time t respectively, and P max is the rated power maximum value of the energy storage device, and η charge , η discharge are the charge and discharge efficiencies of the energy storage system at time t respectively;

[0282] In the power grid supply-demand balance constraint, the power grid needs to meet the load demand in real time. When the load suddenly increases and the energy storage is insufficient, it is necessary to purchase electricity from the power grid at a high price, increasing the energy cost. The energy storage discharge power and the power grid power purchase jointly compensate for the difference between the charging power and the load:

[0283] P discharge (t) - P charge (t) + P grid (t) = D(t);

[0284] In the above formula, P grid (t) is the power grid power purchase at time t, and D(t) is the load demand at time t;

[0285] Device health and maintenance constraints transform predictive maintenance results into actionable maintenance strategies. For example, if the predicted health status of a battery pack indicates that it may fall below the threshold in the next 3 days, maintenance is scheduled during a low-load period at night to minimize the impact on the power grid, including:

[0286] When the predicted health status value is lower than the threshold, maintenance is deferred to avoid immediate downtime that affects the power supply health status. The health status threshold trigger for maintenance constraints is:

[0287]

[0288] When the predicted remaining useful life value is less than the minimum threshold, new equipment is ordered in advance to ensure seamless replacement. The remaining life threshold trigger for replacement constraints is:

[0289]

[0290] In the above formula, SOH(t) is the predicted health status value at time t, SOH critical is the health status threshold, t maintenance is the maintenance plan time point, Δt delay is the delay time, RUL(t) is the predicted remaining useful life value at time t, RUL min is the minimum threshold for the remaining useful life, t replace is the energy storage device replacement plan, Δt replace is the advance order time;

[0291] A mathematical model for the energy storage system configuration optimization model is established. The mixed-integer linear programming (MILP) model is selected to transform the non-linear problem into a linear form:

[0292]

[0293] In the above formula, T horizon is the total time;

[0294] The mathematical model of the energy storage system configuration optimization model simplifies complex constraints through MILP linearization, facilitating rapid solution by commercial solvers. The genetic algorithm or model predictive control (MPC) is used to improve the solution speed. In the real-time bidding scenario of the power market, MPC is used to quickly adjust the charge and discharge plan. In long-term capacity planning, the accurate MILP solution is preferred;

[0295] Rolling horizon optimization is adopted. Every ΔT, the energy storage system configuration optimization model is re-solved based on the latest load prediction and predictive maintenance prediction data. Every ΔT, the trained load prediction model and predictive maintenance model are called to generate the prediction results for the future optimization window, which is used to balance the computational complexity and decision-making foresight.

[0296] Example 2:

[0297] As Figure 2 shown, an energy storage optimization configuration system based on data-driven model fusion includes a data fusion module, a model construction module, a model training module, and a configuration optimization module;

[0298] The data fusion module is used to collect the load data and operation and maintenance data of the energy storage system, and respectively extract the characteristic data that has an important impact on energy storage load prediction and equipment status prediction for data fusion;

[0299] The model construction module is used to construct a load prediction model for predicting the load status of the energy storage equipment based on the fused energy storage load prediction characteristic data; and construct a predictive maintenance model for predicting the health status and remaining service life status of the energy storage equipment based on the fused equipment status prediction characteristic data;

[0300] The model training module is used to iteratively train the load prediction model and the predictive maintenance model of the energy storage system to obtain the trained load prediction model and predictive maintenance model;

[0301] The configuration optimization module is used to utilize the trained load prediction model and predictive maintenance model to predict the load status, health status and remaining service life status of the energy storage equipment in real time, construct an energy storage system configuration optimization model, and optimize the overall configuration of the energy storage system.

[0302] The model construction module includes a load prediction model construction unit;

[0303] The load prediction model construction unit includes an information calculation subunit, an input update subunit, and a hidden state output calculation subunit;

[0304] The information calculation subunit is used to calculate the retained information and new information in the load prediction model based on the fused energy storage load prediction characteristic data by using the following formula:

[0305] F t = σ(W f · [H t-1 , X t + b f );

[0306] I t = σ(W i · [H t-1 , X t + b i );

[0307]

[0308] In the above formula, F tThe information retained in the load prediction model at the current moment, σ is the Sigmoid function, W f is the weight matrix of the retained information, [H t-1 , X t is the concatenation of the hidden state at the previous moment and the input data at the current moment, b f is the bias term of the retained information, I t is the new information in the load prediction model at the current moment, W i is the weight matrix of the new information, b i is the bias term of the new information, is the candidate value vector containing the new information at the current moment, tanh is the hyperbolic tangent function, W c is the weight matrix of the candidate value vector, b c is the bias term of the candidate value vector;

[0309] The input update subunit is used to combine the retained information and the new information in the load prediction model, and update the input of the energy storage system load prediction model using the following formula:

[0310]

[0311] In the above formula, C t is the input of the updated energy storage system load prediction model at the current moment, C t-1 is the input of the updated energy storage system load prediction model at the previous moment;

[0312] The hidden state output calculation subunit is used to calculate the hidden state output at the current moment based on the input of the updated energy storage system load prediction model using the following formula:

[0313] O t = σ(W o · [H t-1 , X t + b o );

[0314] H t = O t · tanh(C t );

[0315] In the above formula, O t is the output information at the current moment, W o is the weight matrix of the output information, b o is the bias term of the output information, H t is the hidden state output at the current moment.

[0316] The model construction module further includes a predictive maintenance model construction unit;

[0317] The predictive maintenance model construction unit includes a self-attention weight generation subunit and a feature matrix convolution subunit;

[0318] The self-attention weight generation subunit is used to capture the spatial and temporal dependence relationships between device state prediction feature data, and generate spatial self-attention weights and temporal self-attention weights:

[0319] α = softmax(S);

[0320] β = sofmtax(U);

[0321]

[0322] In the above formula, α is the spatial self-attention weight, S is the spatial dependence relationship between feature data, β is the temporal self-attention weight, U is the temporal dependence relationship between feature data, Vs is the input feature matrix of spatial self-attention, Ws and Rs are both learnable weight matrices for capturing spatial feature dependence relationships, Vt is the input feature matrix of temporal self-attention, Wt and Rt are both independent weight matrices for capturing temporal feature dependence relationships;

[0323] The feature matrix convolution subunit is used to further extract the spatial and temporal dynamic relationships between feature data by convolution, and obtain the corresponding convolved feature matrix:

[0324] V′ I = A′ * (αV);

[0325]

[0326] In the above formula, V′ I is the spatially convolved feature matrix, A′ is the adjacency matrix after self-attention weighting, V is the input feature matrix, is the temporally convolved feature matrix, Mish is the activation function, W H is the weight matrix of temporal convolution, * is the convolution operation.

[0327] The model training module includes a loss function construction unit, a gradient calculation unit, a model parameter update unit, and an iterative training unit;

[0328] The loss function construction unit is used to consider the characteristics of the energy storage system and the specific requirements of the prediction task, weight different samples, and construct a weighted loss function for the load prediction model and the predictive maintenance model of the energy storage system;

[0329] The weighted loss function of the load prediction model is:

[0330]

[0331] In the above formula, L δ (y1, f(x)1) is the loss function of the load prediction model, M is the total number of load prediction samples, w i is the weight of the i-th load prediction sample, y i，1 is the true value of the i-th load prediction sample, f(x i )1 is the load prediction value of the i-th load prediction sample, and δ is a hyperparameter used to control the sensitivity of the loss function to outliers;

[0332] The weighted loss function of the predictive maintenance model is:

[0333]

[0334] In the above formula, L τ (y2, f(x)2) is the loss function of the predictive maintenance model, N is the total number of predictive maintenance samples, w j is the weight of the j-th predictive maintenance sample, y j，2 is the true value of the j-th predictive maintenance sample, f(x j )2 is the predicted value of the j-th predictive maintenance sample, and τ is a hyperparameter of the quantile;

[0335] The gradient calculation unit is used to randomly select a small batch of samples and calculate the gradients of the loss functions of the load prediction model and the predictive maintenance model;

[0336] The gradient of the loss function of the load prediction model is:

[0337]

[0338] In the above formula, is the gradient of the loss function of the load prediction model with respect to the current model parameters, is the gradient of the load prediction value f(x)1 with respect to the model parameter θt;

[0339] The gradient of the loss function of the predictive maintenance model is:

[0340]

[0341] In the above formula, is the gradient of the loss function of the predictive maintenance model with respect to the current model parameters, is the gradient of the load prediction value f(x)2 with respect to the model parameter ζt;

[0342] The model parameter update unit is used to update the model parameters according to the gradients of the loss functions of the load prediction model and the predictive maintenance model with respect to the corresponding model parameters, using the following formula:

[0343]

[0344] In the above formula, θ t+1 is the parameter of the updated load prediction model, η is the learning rate, which is used to control the step size of parameter update, and ζ t+1 is the parameter of the updated predictive maintenance model

[0345] The iterative training unit is used to return to the loss function construction unit, and repeatedly iteratively calculate the loss function and the corresponding gradient to update the model parameters until the model parameters converge or reach the preset number of iterations, and the training of the load prediction model and the predictive maintenance model ends.

[0346] In the configuration optimization module, the objective function of the energy storage system configuration optimization model includes:

[0347] min(C energy +C maintenance +C replacement +C penalty );

[0348] C energy =λ(t)P grid (t)Δt;

[0349]

[0350] C penalty =γ|D(t)-P discharge (t)|;

[0351] In the above formula, C energy is the energy cost, C maintenance is the maintenance cost, C replacem ent is the equipment replacement cost, C penalty is the penalty cost, λ(t) is the time-of-use electricity price at time t, P grid (t) is the electricity purchased from the power grid at time t, Δt is the time variation, α and β are the fixed cost coefficients of the maintenance cost and the replacement cost respectively, is the maintenance indicator function, which is 1 when a maintenance action is performed and 0 otherwise, is the replacement indicator function, which is 1 when a replacement action is performed and 0 otherwise, γ is the supply-demand imbalance penalty coefficient, D(t) is the load demand at time t, P discharge (t) is the discharge power of the energy storage system at time t;

[0352] The constraint conditions of the energy storage system configuration optimization model include the energy storage physical limit constraint, the power grid supply-demand balance constraint, and the equipment health and maintenance constraint;

[0353] The energy storage physical limit constraint includes:

[0354] 0 ≤ SOC(t) ≤ C storage ;

[0355] 0 ≤ P charge (t) ≤ P max ;

[0356] 0 ≤ P discharge (t) ≤ P max ;

[0357]

[0358] In the above formula, SOC(t) is the state of charge of the energy storage system at time t, and C storage is the energy storage capacity. According to the long-term load prediction results, the energy storage capacity is dynamically adjusted. P charge (t) and P discharge (t) are the charging and discharging powers of the energy storage system at time t respectively, and P max is the maximum rated power of the energy storage device. η charge and η discharge are the charging and discharging efficiencies of the energy storage system at time t respectively;

[0359] The power grid supply-demand balance constraint is:

[0360] P discharge (t) - P charge (t) + P grid (t) = D(t);

[0361] In the above formula, P grid (t) is the power grid power purchase at time t, and D(t) is the load demand at time t;

[0362] The equipment health and maintenance constraints include:

[0363]

[0364] In the above formula, SOH(t) is the predicted value of the health state at time t, and SOH critical is the health state threshold. t maintenance is the maintenance plan time point. When the predicted result of the health state is lower than the threshold, the energy storage device is arranged to stop for maintenance. Δt delay is the delay time, and RUL(t) is the predicted value of the remaining useful life at time t. RUL min is the minimum threshold of the remaining useful life. t replace is the energy storage device replacement plan. When the predicted result of the remaining useful life is close to the end of the life, the equipment replacement process is triggered. Δt replace is the advance order time.

Claims

1. A method for optimizing the configuration of energy storage based on data-driven model fusion, characterized in that the method includes: S1. Collect the load data and operation and maintenance data of the energy storage system, and respectively extract the characteristic data that has an important impact on energy storage load prediction and equipment status prediction for data fusion; S2. Based on the fused characteristic data of energy storage load prediction, construct a load prediction model for predicting the load status of energy storage equipment; based on the fused characteristic data of equipment status prediction, construct a predictive maintenance model for predicting the health status and remaining service life status of energy storage equipment; S3. Iteratively train the load prediction model and the predictive maintenance model of the energy storage system to obtain the trained load prediction model and predictive maintenance model; S4. Use the trained load prediction model and predictive maintenance model to predict the load status, health status and remaining service life status of energy storage equipment in real time, construct an energy storage system configuration optimization model, and optimize the overall configuration of the energy storage system.

2. The method for optimizing the configuration of energy storage based on data-driven model fusion according to claim 1, characterized in that in S2, the specific steps for constructing the load prediction model include: A. Based on the fused characteristic data of energy storage load prediction, use the following formula to calculate the retained information and new information in the load prediction model: F t = σ(W f · [H t-1 , X t + b f ); I t = σ(W i · [H t-1 , X t + b i ); In the above formula, F t is the retained information in the load prediction model at the current moment, σ is the Sigmoid function, W f is the weight matrix of the retained information, [H t-1 , X t is the concatenation of the hidden state at the previous moment and the input data at the current moment, b f is the bias term of the retained information, I t is the new information in the load prediction model at the current moment, W i is the weight matrix of the new information, b i is the bias term of the new information, is the candidate value vector containing the new information at the current moment, tanh is the hyperbolic tangent function, W c is the weight matrix of the candidate value vector, b c is the bias term of the candidate value vector; B. Combine the retained information and new information in the load prediction model, and use the following formula to update the input of the energy storage system load prediction model: In the above formula, C t is the input of the energy storage system load prediction model updated at the current moment, and C t-1 is the input of the energy storage system load prediction model updated at the previous moment; C. Based on the updated input of the energy storage system load prediction model, use the following formula to calculate the hidden state output at the current moment: O t = σ(W o · [H t-1 , X t + b O ); H t = O t ·tanh(C t ); In the above formula, O t is the output information at the current moment, W o is the weight matrix of the output information, b o is the bias term of the output information, and H t is the hidden state output at the current moment.

3. The method for optimizing the configuration of energy storage based on data-driven model fusion according to claim 1, characterized in that in S2, the specific steps for constructing the predictive maintenance model include: a. Capture the spatial and temporal dependence relationships between the characteristic data of equipment status prediction, and generate spatial self-attention weights and temporal self-attention weights: α = softmax(S); β = softmax(U); S = (V s Q s W s T )R s W s (V s Q s W s T ); U=(V t Q t W t T )W t (V t Q t W t T )R t ; In the above formula, α is the spatial self-attention weight, S is the spatial dependence relationship between feature data, β is the temporal self-attention weight, U is the temporal dependence relationship between feature data, and V s is the input feature matrix of spatial self-attention, Q s , W s , R s are all learnable weight matrices for capturing spatial feature dependence relationships, and V t is the input feature matrix of temporal self-attention, Q t , W t , R t are all independent weight matrices for capturing temporal feature dependence relationships; b. Use convolution to further extract the spatial and temporal dynamic relationships between the characteristic data to obtain the corresponding convolved characteristic matrix: V I = A′★(αV); V Q = Mish(W H ★(βV I )); In the above formula, V′ I is the feature matrix after spatial convolution, A′ is the adjacency matrix after self-attention weighting, V is the input feature matrix, V Q is the feature matrix after temporal convolution, Mish is the activation function, W H is the weight matrix of temporal convolution, and ★ is the convolution operation.

4. The method for optimizing the configuration of energy storage based on data-driven model fusion according to claim 1, characterized in that S3 includes: S31. Considering the characteristics of the energy storage system and the specific requirements of the prediction task, weight different samples to construct a weighted loss function for the load prediction model and the predictive maintenance model of the energy storage system; The weighted loss function of the load prediction model is: In the above formula, L δ (y1, f(x)1) is the loss function of the load prediction model, M is the total number of load prediction samples, w i is the weight of the load prediction sample i, y i，1 is the true value of the load prediction sample i, f(x i )1 is the load prediction value of the load prediction sample i, and δ is a hyperparameter used to control the sensitivity of the loss function to outliers; The weighted loss function of the predictive maintenance model is: In the above formula, L τ (y2, f(x)2) is the loss function of the predictive maintenance model, N is the total number of predictive maintenance samples, w j is the weight of the predictive maintenance sample j, y j，2 is the true value of the predictive maintenance sample j, f(x j )2 is the predicted value of the predictive maintenance sample j, and τ is the hyperparameter of the quantile; S32. Randomly select a small batch of samples, and calculate the gradients of the loss functions of the load prediction model and the predictive maintenance model; The gradient of the loss function of the load prediction model is: In the above formula, is the gradient of the loss function of the load prediction model with respect to the current model parameters, is the gradient of the load prediction value f(x)1 with respect to the model parameter θ t ; The gradient of the loss function of the predictive maintenance model is: In the above formula, is the gradient of the loss function of the predictive maintenance model with respect to the current model parameters, is the gradient of the load prediction value f(x)2 with respect to the model parameter ζ t ; S33. According to the gradients of the loss functions of the load prediction model and the predictive maintenance model with respect to the corresponding model parameters, use the following formula to update the model parameters: In the above formula, θ t+1 is the updated load prediction model parameter, η is the learning rate, which is used to control the step size of parameter update, and ζ t+1 is the updated predictive maintenance model parameter; S34. Return to step S31, and repeatedly perform iterative calculations of the loss function and the corresponding gradients to update the model parameters until the model parameters converge or reach the preset number of iterations, at which point the training of the load prediction model and the predictive maintenance model ends.

5. A method for optimizing the configuration of an energy storage system based on data-driven model fusion according to claim 1, wherein in S4, the objective function of the energy storage system configuration optimization model includes: min(C energy +C maintenance +C replacement +C penalty ); C energy = λ(t)P grid (t)Δt; C penalty = γ |D(t) - P discharge (t)|; In the above formula, C energy is the energy cost, C maintenance is the maintenance cost, C replacement is the equipment replacement cost, C penalty is the penalty cost, λ(t) is the time-of-use electricity price at time t, P grid (t) is the electricity purchased from the power grid at time t, △t is the time variation, α and β are the fixed cost coefficients of the maintenance cost and the replacement cost respectively, is the maintenance indicator function, which is 1 when the maintenance action is performed and 0 otherwise, is the replacement indicator function, which is 1 when the replacement action is performed and 0 otherwise, γ is the penalty coefficient for supply-demand imbalance, D(t) is the load demand at time t, P discharge (t) is the discharge power of the energy storage system at time t; The constraint conditions of the energy storage system configuration optimization model include energy storage physical limit constraints, power grid supply-demand balance constraints, and equipment health and maintenance constraints; The energy storage physical limit constraints include: 0 ≤ SOC(t) ≤ C storage ; 0 ≤ P charge (t) ≤ P max ; 0 ≤ P discharge (t) ≤ P max ; In the above formula, SOC(t) is the state of charge of the energy storage system at time t, and C storage is the energy storage capacity. According to the long-term load prediction results, the energy storage capacity is dynamically adjusted. P charge (t) and P discharge (t) are the charging and discharging powers of the energy storage system at time t, respectively. P max is the maximum rated power of the energy storage device. η charge and η discharge are the charging and discharging efficiencies of the energy storage system at time t, respectively; The power grid supply-demand balance constraint is: P discharge (t) - P charge (t) + P grid (t) = D(t); In the above formula, P grid (t) is the electricity purchased from the power grid at time t, and D(t) is the load demand at time t; The equipment health and maintenance constraints include: In the above formula, SOH(t) is the predicted value of the health state at time t, and SOH critical is the health state threshold, t maintenance is the maintenance plan time point. When the predicted result of the health state is lower than the threshold, arrange for the energy storage device to stop for maintenance. Δt delay is the delay time, and RUL(t) is the predicted value of the remaining useful life at time t. RUL min is the minimum threshold of the remaining useful life, t replace is the energy storage device replacement plan. When the predicted result of the remaining useful life approaches the end of life, trigger the device replacement process. Δt replace is the advance order time.

6. An energy storage optimization configuration system based on data-driven model fusion, wherein the system includes a data fusion module, a model construction module, a model training module, and a configuration optimization module; The data fusion module is used to collect the load data and operation and maintenance data of the energy storage system, and respectively extract the characteristic data that has an important impact on energy storage load prediction and equipment status prediction for data fusion; The model construction module is used to construct a load prediction model for predicting the load status of energy storage equipment based on the fused energy storage load prediction characteristic data; and construct a predictive maintenance model for predicting the health status and remaining service life status of energy storage equipment based on the fused equipment status prediction characteristic data; The model training module is used to iteratively train the load prediction model and the predictive maintenance model of the energy storage system to obtain the trained load prediction model and predictive maintenance model; The configuration optimization module is used to use the trained load prediction model and predictive maintenance model to predict the load status, health status, and remaining service life status of energy storage equipment in real time, construct an energy storage system configuration optimization model, and optimize the overall configuration of the energy storage system.

7. A method for optimizing the configuration of an energy storage system based on data-driven model fusion according to claim 6, wherein the model construction module includes a load prediction model construction unit; The load prediction model construction unit includes an information calculation subunit, an input update subunit, and a hidden state output calculation subunit; The information calculation subunit is used to calculate the retained information and new information in the load prediction model based on the fused energy storage load prediction characteristic data using the following formula: F t = σ(W f · [H t-1 , X t + b f ) ; I t = σ(W i · [H t-1 , X t + b i ); In the above formula, F t is the retained information in the load prediction model at the current moment, σ is the Sigmoid function, and W f is the weight matrix of the retained information. [H t-1 , X t is the concatenation of the hidden state at the previous moment and the input data at the current moment. b f is the bias term of the retained information. I t is the new information in the load prediction model at the current moment. W i is the weight matrix of the new information. b i is the bias term of the new information. is the candidate value vector containing the new information at the current moment. tanh is the hyperbolic tangent function. W c is the weight matrix of the candidate value vector. b c is the bias term of the candidate value vector; The input update subunit is used to update the input of the energy storage system load prediction model by combining the retained information and new information in the load prediction model using the following formula: In the above formula, C t is the input of the energy storage system load prediction model updated at the current moment, and C t-1 is the input of the energy storage system load prediction model updated at the previous moment; The hidden state output calculation subunit is used to calculate the hidden state output at the current moment based on the updated input of the energy storage system load prediction model using the following formula: O t = σ(W o · [H t-1 , X t + b o ); H t = O t ·tanh(C t ); In the above formula, O t is the output information at the current moment, W o is the weight matrix of the output information, b o is the bias term of the output information, and H t is the hidden state output at the current moment.

8. A method for optimizing the configuration of an energy storage system based on data-driven model fusion according to claim 6, wherein the model construction module further includes a predictive maintenance model construction unit; The predictive maintenance model construction unit includes a self-attention weight generation subunit and a feature matrix convolution subunit; The self-attention weight generation subunit is used to capture the spatial and temporal dependencies between the device state prediction feature data, and generate spatial self-attention weights and temporal self-attention weights: α = softmax(S); β = softmax(U); S = (V s Q s W s T )R s W s (V s Q s W s T ); U = (V t Q t W t T )W t (V t Q t W t T )R t ; In the above formula, α is the spatial self-attention weight, S is the spatial dependence relationship between feature data, β is the temporal self-attention weight, U is the temporal dependence relationship between feature data, and V s is the input feature matrix of spatial self-attention, Q s , W s , R s are all learnable weight matrices for capturing spatial feature dependence relationships, and V t is the input feature matrix of temporal self-attention, Q t , W t , R t are all independent weight matrices for capturing temporal feature dependence relationships; The feature matrix convolution subunit is used to further extract the spatial and temporal dynamic relationships between the feature data by convolution, and obtain the corresponding convolved feature matrix: V′ I = A′ * (αV); V Q = Mish(V H ★(βV I )); In the above formula, V′ I is the feature matrix after spatial convolution, A′ is the adjacency matrix after self-attention weighting, V is the input feature matrix, V Q is the feature matrix after temporal convolution, Mish is the activation function, W H is the weight matrix of temporal convolution, and ★ is the convolution operation.

9. The energy storage optimization configuration system based on data-driven model fusion according to claim 6, characterized in that The model training module includes a loss function construction unit, a gradient calculation unit, a model parameter update unit, and an iterative training unit; The loss function construction unit is used to consider the characteristics of the energy storage system and the specific requirements of the prediction task, weight different samples, and construct weighted loss functions for the load prediction model and the predictive maintenance model of the energy storage system; The weighted loss function of the load prediction model is: In the above formula, L δ (y1, f(x)1) is the loss function of the load prediction model, M is the total number of load prediction samples, w i is the weight of the load prediction sample i, y i，1 is the true value of the load prediction sample i, f(x i )1 is the load prediction value of the load prediction sample i, and δ is a hyperparameter used to control the sensitivity of the loss function to outliers; The weighted loss function of the predictive maintenance model is: In the above formula, L τ (y2, f(x)2) is the loss function of the predictive maintenance model, N is the total number of predictive maintenance samples, w j is the weight of the predictive maintenance sample j, y j，2 is the true value of the predictive maintenance sample j, f(x j )2 is the predicted value of the predictive maintenance sample j, and τ is the hyperparameter of the quantile; The gradient calculation unit is used to randomly select a small batch of samples and calculate the gradients of the loss functions of the load prediction model and the predictive maintenance model; The gradient of the loss function of the load prediction model is: In the above formula, is the gradient of the loss function of the load prediction model with respect to the current model parameters, is the gradient of the load prediction value f(x)1 with respect to the model parameter θ t ; The gradient of the loss function of the predictive maintenance model is: In the above formula, is the gradient of the predictive maintenance model loss function with respect to the current model parameters, is the gradient of the load prediction value f(x)2 with respect to the model parameter ζ t ; The model parameter update unit is used to update the model parameters according to the gradients of the loss functions of the load prediction model and the predictive maintenance model with respect to the corresponding model parameters, using the following formula: In the above formula, θ t+1 is the updated load prediction model parameter, η is the learning rate, which is used to control the step size of parameter update, and ζ t+1 is the updated predictive maintenance model parameter The iterative training unit is used to return to the loss function construction unit, repeatedly iterate to calculate the loss function and the corresponding gradients to update the model parameters until the model parameters converge or reach the preset number of iterations, and the training of the load prediction model and the predictive maintenance model ends.

10. The energy storage optimization configuration system based on data-driven model fusion according to claim 6, characterized in that In the configuration optimization module, the objective function of the energy storage system configuration optimization model includes: min(C energy +C maintenance +C replacement +C penalty ); C energy = λ(t)P grid (t)Δt; C penalty = γ|D(t) - P discharge (t)|; In the above formula, C energy is the energy cost, C maintenance is the maintenance cost, C replacement is the equipment replacement cost, C penalty is the penalty cost, λ(t) is the time-of-use electricity price at time t, P grid (t) is the electricity purchased from the power grid at time t, Δt is the time variation, α and β are the fixed cost coefficients of the maintenance cost and the replacement cost respectively, is the maintenance indicator function, which is 1 when the maintenance action is carried out and 0 otherwise, is the replacement indicator function, which is 1 when the replacement action is carried out and 0 otherwise, γ is the penalty coefficient for supply-demand imbalance, D(t) is the load demand at time t, P discharge (t) is the discharge power of the energy storage system at time t; The constraint conditions of the energy storage system configuration optimization model include energy storage physical limit constraints, power grid supply-demand balance constraints, and equipment health and maintenance constraints; The energy storage physical limit constraints include: 0 ≤ SOC(t) ≤ C storage ; 0 ≤ P charge (t) ≤ P max ; 0 ≤ P discharge (t) ≤ P max ; In the above formula, SOC(t) is the state of charge of the energy storage system at time t, and C storage is the energy storage capacity. According to the long-term load prediction results, the energy storage capacity is dynamically adjusted. P charge (t) and P discharge (t) are the charging and discharging powers of the energy storage system at time t respectively. P max is the maximum rated power of the energy storage device. η charge and η discharge are the charging and discharging efficiencies of the energy storage system at time t respectively; The power grid supply-demand balance constraint is: P discharge (t) - P charge (t) + P grid (t) = D(t); In the above formula, P grid (t) represents the electricity purchased from the power grid at time t, and D(t) represents the load demand at time t; The equipment health and maintenance constraints include: In the above formula, SOH(t) is the predicted value of the health state at time t, and SOH critical is the health state threshold, and t maintenance is the maintenance plan time point. When the predicted result of the health state is lower than the threshold, the energy storage device is arranged for shutdown maintenance, and Δt delay is the delay time, and RUL(t) is the predicted value of the remaining useful life at time t, and RUL min is the minimum threshold of the remaining useful life, and t replace is the replacement plan of the energy storage device. When the predicted result of the remaining useful life is close to the end of life, the device replacement process is triggered, and Δt replace is the advance ordering time.

Images