Intelligent optimization method and system for revenue mode of commercial energy storage power station

Abstract

The invention provides an intelligent optimization method and system for a revenue mode of a commercial energy storage power station, and relates to the technical field of energy storage power stations. Establishing a multi-dimensional life loss coefficient matrix by analyzing the health state data of the energy storage equipment; a candidate charging and discharging scheduling strategy is generated based on a deep reinforcement learning model, and a double-constraint optimization charging and discharging strategy with predicted income and life loss as a reward function is adopted; and finally generating a control instruction to realize optimized operation of the energy storage power station. According to the method, equipment life loss and economic benefits are comprehensively considered, and long-term benefit maximization of the energy storage power station is realized.

Description

Technical Field

[0001] The present invention relates to energy storage power station technology, and in particular to a method and system for intelligently optimizing the revenue model of a commercial energy storage power station. Background Art

[0002] As the scale of renewable energy integration continues to expand, the grid's demand for flexible resources is growing. Commercial energy storage power plants, as a key source of power system flexibility, play a vital role in peak and frequency regulation, capacity support, and power quality improvement. Commercial energy storage power plants generate economic benefits by charging during low-price periods, discharging during high-price periods, participating in the power ancillary services market, and providing demand response services. However, the profitability of energy storage power plants is directly related to their operating strategies. Maximizing economic benefits while ensuring equipment safety and extending service life is a key issue in energy storage plant operations.

[0003] Currently, commercial energy storage power station revenue model optimization is primarily based on traditional mathematical programming methods, such as linear programming, dynamic programming, and heuristic algorithms. These methods have the following flaws and shortcomings when faced with complex factors such as real-time grid electricity price fluctuations, energy storage equipment lifespan loss, and volatile market environments: Existing energy storage power station scheduling optimization methods generally lack accurate assessment of battery life loss. They often use simplified linear models, which cannot accurately reflect the comprehensive impact of multi-dimensional factors such as cycle number, deep cycle rate and operating temperature on battery life. As a result, while pursuing short-term economic benefits, excessive loss of battery life is achieved, reducing the long-term economic efficiency of the energy storage system.

[0004] Traditional optimization algorithms perform poorly when dealing with uncertainties such as real-time fluctuations in grid electricity prices. They typically formulate fixed scheduling strategies based on historical data and simple forecasting models, making it difficult to adapt to real-time changes in the electricity market and unable to make flexible adjustments in the event of drastic price fluctuations or emergencies. This means that the profit potential of energy storage power stations has not been fully tapped.

[0005] The profit optimization models in existing technologies mostly adopt single-objective optimization, either focusing too much on short-term economic benefits and ignoring equipment life, or being overly conservative to extend equipment service life at the expense of economic benefits. They lack a comprehensive optimization mechanism that uses maximizing profits and minimizing life loss as dual constraints, and cannot achieve the optimal economy throughout the life cycle of the energy storage system. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for intelligently optimizing the revenue model of a commercial energy storage power station, which can solve the problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for intelligently optimizing a revenue model of a commercial energy storage power station, comprising: Acquire historical operating data of the commercial energy storage power station, the historical operating data including charge and discharge power data, battery state of charge data, real-time grid electricity price data, and energy storage equipment health status data; Building a deep reinforcement learning model for a commercial energy storage power station based on the historical operating data, wherein the deep reinforcement learning model uses a deep neural network structure to nonlinearly map the historical operating data through the deep neural network structure to generate a predicted profit value; Analyzing the health status data of the energy storage device through the deep reinforcement learning model, establishing a multi-dimensional life loss coefficient matrix based on the number of cycles, deep cycle rate, and operating temperature, and realizing the assessment of the equipment life loss; Based on the deep reinforcement learning model, multiple groups of candidate charge and discharge scheduling strategies are generated. The predicted revenue and life loss are used as dual constraints of the reward function. By alternately training the policy network and the value network, the generation of charge and discharge strategies that maximize revenue while considering the life of the equipment is achieved. An operation control instruction is generated according to the optimized charge and discharge scheduling strategy, and the operation control instruction is sent to the energy management system of the commercial energy storage power station to control the energy storage power station to operate according to the optimized charge and discharge scheduling strategy.

[0008] A deep reinforcement learning model for a commercial energy storage power station is constructed based on the historical operating data. The deep reinforcement learning model adopts a deep neural network structure. The historical operating data is nonlinearly mapped by the deep neural network structure to generate a predicted profit value. The generated predicted profit value includes: The deep reinforcement learning model adopts a multi-layer feedforward neural network structure, which includes an input layer, a hidden layer, and an output layer, wherein the hidden layer performs feature extraction and nonlinear mapping on the historical operation data through a weight matrix, a bias vector, and a nonlinear activation function; The network parameters of the multi-layer feedforward neural network structure are constructed as random variables, a Gaussian distribution model of the network parameters is established using a variational inference method, and the prediction uncertainty of the multi-layer feedforward neural network structure is quantitatively characterized by the mean parameter and variance parameter of the Gaussian distribution model; Adding disturbances to the historical operating data based on the Gaussian distribution model to generate adversarial samples, and using the adversarial samples in the training process of the multi-layer feedforward neural network structure to improve the robustness of the multi-layer feedforward neural network structure; Calculating a prediction mean and a prediction standard deviation based on the prediction results of the multi-layer feedforward neural network structure, and determining a confidence interval of the prediction results according to the prediction mean and the prediction standard deviation; calculating a dynamic adjustment coefficient according to the prediction standard deviation, and adaptively optimizing the prediction results of the multi-layer feedforward neural network structure based on the dynamic adjustment coefficient to obtain an optimized prediction result; The optimized prediction result and the confidence interval are input into a profit calculation module, and a predicted profit value of the commercial energy storage power station is generated by the profit calculation module.

[0009] A Gaussian distribution model of the network parameters is established using a variational inference method, and the prediction uncertainty of the multi-layer feedforward neural network structure is quantitatively characterized by the mean parameter and variance parameter of the Gaussian distribution model. The method includes: Obtaining training data for a multi-layer feedforward neural network structure, constructing a likelihood function based on the training data, selecting a Gaussian distribution as a prior distribution of network parameters, and combining the likelihood function with the prior distribution using a Bayesian formula to construct a posterior distribution; Approximating the posterior distribution using a variational inference method to construct a variational distribution family in the form of a Gaussian distribution, wherein the variational distribution family includes a mean parameter vector and a variance parameter vector, and obtaining optimal variational distribution parameters by minimizing the Kullback-Leibler divergence between the variational distribution family and the posterior distribution using the variational inference method; Inputting the optimal variational distribution parameters into the multi-layer feedforward neural network structure, and for each layer in the multi-layer feedforward neural network structure, constructing the parameter distribution of the layer based on the mean parameter vector and the variance parameter vector using the variational inference method; obtaining network parameter samples based on the parameter distribution, and inputting the network parameter samples into the multi-layer feedforward neural network structure to obtain corresponding network outputs; Repeatedly obtain multiple groups of network parameter samples, input the multiple groups of network parameter samples into the multi-layer feedforward neural network structure to obtain multiple groups of network outputs, calculate the prediction mean and the prediction variance based on the multiple groups of network outputs; and use the prediction mean and the prediction variance as the prediction uncertainty quantification results of the multi-layer feedforward neural network structure.

[0010] The deep reinforcement learning model is used to analyze the health status data of the energy storage device, and a multi-dimensional life loss coefficient matrix is ​​established based on the number of cycles, deep cycle rate, and operating temperature to achieve an assessment of the device life loss, including: Acquire health status data of the energy storage device, construct a state feature vector based on the health status data, and use the state feature vector as input data for a deep reinforcement learning method; Based on the state feature vector, the deep reinforcement learning method is used to extract the number of cycles, deep cycle rate, and operating temperature parameters of the energy storage device. Based on the deep reinforcement learning method, a time decay factor is introduced to adaptively weight the operating data at different times in the state feature vector to obtain weighted historical loss data; Analyzing the dynamic change characteristics of the number of cycles, the deep cycle rate, and the operating temperature parameters based on the weighted historical loss data and inputting the dynamic change characteristics into a loss benchmark model to obtain a benchmark loss characteristic of the energy storage device; Based on the benchmark loss characteristics, a multi-dimensional life loss coefficient matrix is ​​constructed using the deep reinforcement learning method, and the multi-dimensional life loss coefficient matrix is ​​optimized by minimizing the deviation between the predicted loss value and the actual loss value to obtain an optimized loss coefficient matrix; The optimized loss coefficient matrix is ​​integrated with time to obtain a cumulative loss state of the energy storage device, and the remaining life of the energy storage device is determined based on the cumulative loss state and the design life of the energy storage device.

[0011] The deep reinforcement learning method is used to construct a multi-dimensional life loss coefficient matrix, and the multi-dimensional life loss coefficient matrix is ​​optimized by minimizing the deviation between the predicted loss value and the actual loss value. The optimized loss coefficient matrix includes: A multi-dimensional life loss coefficient matrix includes a cycle number dimension, a deep cycle rate dimension, and a temperature dimension; minimizing the deviation between the predicted loss value output by the multi-dimensional life loss coefficient matrix and the actual loss value; The multi-dimensional life loss coefficient matrix is ​​optimized based on the result of the minimization process to obtain an optimized loss coefficient matrix.

[0012] Based on the deep reinforcement learning model, multiple groups of candidate charge and discharge scheduling strategies are generated. The predicted revenue and life loss are used as dual constraints of the reward function. Through alternating training of the policy network and the value network, the generation of charge and discharge strategies that maximize revenue while considering the life of the equipment is achieved. The generation includes: A strategy generation network is constructed based on a deep reinforcement learning model. The strategy generation network includes a parameter generation module and a value assessment module. The network uses the grid load status, electricity price information, device charge status, and ambient temperature of the energy storage device as the system state to generate corresponding charge and discharge action strategy parameters. The system state and the charge and discharge action strategy parameters are combined to form a state-action pair. constructing a dual-constraint reward function based on the state-action pair, inputting the system state into the parameter generation module to obtain a mean parameter and a standard deviation parameter of the action distribution, and inputting the system state and the dual-constraint reward function into the value evaluation module to obtain a state value; An advantage value is calculated based on the state value and the dual-constraint reward function, the advantage value is used to guide the parameter update of the policy generation network, and the temporal difference error of the state value is used to guide the parameter update of the policy generation network, and an optimized policy generation network is obtained through alternating training; The prediction result of the optimized strategy generation network is substituted into the dual-constraint reward function to obtain a strategy evaluation index. Based on the strategy evaluation index, the optimal charge and discharge strategy is selected under the condition of satisfying the life loss constraint to obtain the optimal charge and discharge control strategy of the energy storage device.

[0013] Substituting the prediction result of the optimized strategy generation network into the dual-constraint reward function to obtain a strategy evaluation index, and selecting the optimal charge and discharge strategy based on the strategy evaluation index while satisfying the life loss constraint includes: Substituting the prediction results of the optimized strategy generation network into a dual-constraint reward function, the dual-constraint reward function including a revenue index and a life loss index of the energy storage device, and evaluating the prediction results based on the dual-constraint reward function to obtain a strategy evaluation index corresponding to the prediction results; Calculating a cumulative life loss value of the energy storage device based on the life loss index in the prediction result, and comparing the cumulative life loss value with a preset loss threshold to obtain a life loss constraint condition of the prediction result; The prediction results that meet the life loss constraint condition are sorted in descending order according to the size of the strategy evaluation index, and the results of the descending sorting are used as candidate charging and discharging strategies for the energy storage device; and the charging and discharging strategy with the largest strategy evaluation index among the candidate charging and discharging strategies is used as the optimal charging and discharging control strategy for the energy storage device.

[0014] A second aspect of an embodiment of the present invention provides an intelligent optimization system for a commercial energy storage power station revenue model, including: The first unit is configured to obtain historical operating data of the commercial energy storage power station, wherein the historical operating data includes charge and discharge power data, battery state of charge data, real-time grid electricity price data, and energy storage device health status data; A second unit is configured to construct a deep reinforcement learning model for a commercial energy storage power station based on the historical operating data, wherein the deep reinforcement learning model adopts a deep neural network structure and nonlinearly maps the historical operating data through the deep neural network structure to generate a predicted profit value; A third unit is configured to analyze the health status data of the energy storage device using the deep reinforcement learning model, establish a multi-dimensional life loss coefficient matrix based on the number of cycles, deep cycle rate, and operating temperature, and implement an assessment of the life loss of the device; A fourth unit is configured to generate multiple sets of candidate charge and discharge scheduling strategies based on the deep reinforcement learning model, using predicted revenue and life loss as dual constraints of the reward function, and to generate charge and discharge strategies that maximize revenue while taking into account equipment life through alternating training of the policy network and the value network; The fifth unit is used to generate an operation control instruction according to the optimized charge and discharge scheduling strategy, and send the operation control instruction to the energy management system of the commercial energy storage power station to control the energy storage power station to operate according to the optimized charge and discharge scheduling strategy.

[0015] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0016] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0017] The beneficial effects of this application are as follows: The intelligent optimization method for the revenue model of a commercial energy storage power station provided by the present invention analyzes and learns historical operating data through deep reinforcement learning technology, which can accurately predict the fluctuation trend of market electricity prices, accurately grasp the charging and discharging timing of the energy storage power station, and improve the ability to obtain economic benefits.

[0018] This method innovatively incorporates equipment life loss into the optimization decision-making process and establishes a multi-dimensional life loss assessment system based on the number of cycles, deep cycle rate and operating temperature. While pursuing short-term economic benefits, it effectively balances the long-term equipment usage costs and extends the service life of the energy storage system.

[0019] The present invention realizes intelligent adaptive adjustment of the operating strategy of commercial energy storage power stations. Through the alternating training mechanism of the strategy network and the value network, the system can automatically optimize the charging and discharging strategy according to the dynamic changes of the power grid environment and equipment status, reducing manual intervention and improving the operating efficiency and reliability of the energy storage power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the process of intelligent optimization method of revenue model of commercial energy storage power station according to an embodiment of the present invention; Figure 2 This is a logic block diagram of the deep reinforcement learning profit prediction method for commercial energy storage power stations according to an embodiment of the present invention; Figure 3 This is a graph showing the convergence of the deep reinforcement learning model and the reward function change trend in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0022] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0023] Figure 1 FIG. 1 is a flow chart of an intelligent optimization method for a commercial energy storage power station revenue model according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire historical operating data of the commercial energy storage power station, the historical operating data including charge and discharge power data, battery state of charge data, real-time grid electricity price data, and energy storage equipment health status data; Building a deep reinforcement learning model for a commercial energy storage power station based on the historical operating data, wherein the deep reinforcement learning model uses a deep neural network structure to nonlinearly map the historical operating data through the deep neural network structure to generate a predicted profit value; Analyzing the health status data of the energy storage device through the deep reinforcement learning model, establishing a multi-dimensional life loss coefficient matrix based on the number of cycles, deep cycle rate, and operating temperature, and realizing the assessment of the equipment life loss; Based on the deep reinforcement learning model, multiple groups of candidate charge and discharge scheduling strategies are generated. The predicted revenue and life loss are used as dual constraints of the reward function. By alternately training the policy network and the value network, the generation of charge and discharge strategies that maximize revenue while considering the life of the equipment is achieved. An operation control instruction is generated according to the optimized charge and discharge scheduling strategy, and the operation control instruction is sent to the energy management system of the commercial energy storage power station to control the energy storage power station to operate according to the optimized charge and discharge scheduling strategy.

[0024] In an optional embodiment, a deep reinforcement learning model for a commercial energy storage power station is constructed based on the historical operating data. The deep reinforcement learning model uses a deep neural network structure. The historical operating data is nonlinearly mapped by the deep neural network structure to generate a predicted profit value, including: The deep reinforcement learning model adopts a multi-layer feedforward neural network structure, which includes an input layer, a hidden layer, and an output layer, wherein the hidden layer performs feature extraction and nonlinear mapping on the historical operation data through a weight matrix, a bias vector, and a nonlinear activation function; The network parameters of the multi-layer feedforward neural network structure are constructed as random variables, a Gaussian distribution model of the network parameters is established using a variational inference method, and the prediction uncertainty of the multi-layer feedforward neural network structure is quantitatively characterized by the mean parameter and variance parameter of the Gaussian distribution model; Adding disturbances to the historical operating data based on the Gaussian distribution model to generate adversarial samples, and using the adversarial samples in the training process of the multi-layer feedforward neural network structure to improve the robustness of the multi-layer feedforward neural network structure; Calculating a prediction mean and a prediction standard deviation based on the prediction results of the multi-layer feedforward neural network structure, and determining a confidence interval of the prediction results according to the prediction mean and the prediction standard deviation; calculating a dynamic adjustment coefficient according to the prediction standard deviation, and adaptively optimizing the prediction results of the multi-layer feedforward neural network structure based on the dynamic adjustment coefficient to obtain an optimized prediction result; The optimized prediction result and the confidence interval are input into a profit calculation module, and a predicted profit value of the commercial energy storage power station is generated by the profit calculation module.

[0025] like Figure 2 As shown, the method includes: Historical operation data collection involves multiple dimensions and is divided into two categories: real-time data and historical data. Real-time data includes: grid frequency (sampling frequency 1Hz), battery terminal voltage (sampling frequency 0.1Hz), battery current (sampling frequency 0.1Hz), single cell voltage (sampling frequency 0.01Hz), battery surface temperature (sampling frequency 0.01Hz), and ambient temperature (sampling frequency 0.01Hz). Historical data includes: peak, flat, and valley time-of-use electricity price data (updated daily), historical charge and discharge power curves (15-minute intervals), historical SOC change curves (15-minute intervals), and historical capacity change data (tested monthly).

[0026] Missing value processing (linear interpolation fills single-point missing values, and forward filling handles continuous missing values), outlier processing (3σ principle filters out obvious outliers), normalization (Min-Max standardization scales each feature to the range [0, 1]), and time series feature extraction (sliding window method extracts 24-hour periodic features). After preprocessing, the feature matrix dimensions are (number of samples × 24 × number of features), with 8 features, including hourly electricity price, SOC value, charging power, discharging power, ambient temperature, battery temperature, grid frequency, and day type (weekday / holiday).

[0027] The number of input layer nodes is 24×8=192, corresponding to 8 features in 24 hours; the number of nodes in the first hidden layer is 128, the ReLU activation function is used, the weight matrix dimension is 192×128, and the initialization method is He initialization (the weight value obeys the mean 0 and the standard deviation is sqrt(2 / fan in ) normal distribution), the bias vector is initialized to 0; the second hidden layer has 256 nodes, uses the ReLU activation function, and the weight matrix dimension is 128×256. It is initialized the same as the first layer, and adds a Batch Normalization layer with parameters including the scaling factor γ (initial value is 1) and the offset factor β (initial value is 0); the third hidden layer has 128 nodes, uses the ReLU activation function, and the weight matrix dimension is 256×128. It is initialized the same as before, and adds a Dropout layer with a dropout rate of 0.3; the output layer has 24 nodes, uses a linear activation function, and the weight matrix dimension is 128×24.

[0028] The total number of parameters is (192×128)+(128×256)+(256×128)+(128×24)+128+256+128+24=93720. The forward propagation calculation order is: input data--linear transformation of the first hidden layer--ReLU activation--linear transformation of the second hidden layer--Batch Normalization--ReLU activation--Dropout--linear transformation of the third hidden layer--ReLU activation--linear transformation of the output layer--output result.

[0029] The variational inference method uses the reparameterization technique, where the network parameters are no longer deterministic values ​​but probability distributions. For each weight w, its variational posterior distribution is defined as mean μ and variance σ 2 The specific implementation process is to create two trainable variables μ and ρ for each weight parameter w, where μ directly corresponds to the distribution mean, and σ is obtained by transforming σ = log(1 + exp(ρ)) to ensure that the variance is always positive. When sampling, noise ε is sampled from a standard normal distribution, and weight samples are obtained through w = μ + σ × ε.

[0030] During backpropagation, the gradient can be directly transferred to μ and ρ through this transformation. The training objective function consists of two parts: the negative log-likelihood (prediction error) and the KL divergence regularization term. The ratio of these two components is controlled by the hyperparameter λ, which is typically set to the inverse of the number of samples. Optimization uses the Adam optimizer with an initial learning rate of 0.001, β1=0.9, β2=0.999, ε=1e-8, a weight decay coefficient of 1e-4, and a learning rate decay strategy of 0.9 times the original value every 10 epochs. The training process is iterative in batches, with 64 samples randomly sampled in each batch. The forward propagation, loss function, and gradients are calculated to update the μ and ρ parameters.

[0031] The parameter calculation of the Gaussian distribution model involves precise statistical methods. For a trained model, the distribution of each weight parameter w is composed of mean μ and variance σ. 2 During the prediction phase, Monte Carlo sampling is performed T = 100 times, each time sampling a complete set of network parameters from the posterior distribution of all weights, and performing forward propagation to obtain a prediction result.

[0032] For each weight parameter w, from its Gaussian distribution N(μ,σ 2 ) in the sample value w (t) =μ+σ×ε (t) , where ε (t) is the standard normal random number sampled for the tth time; all sampled weights are combined into a complete set of network parameters, and forward propagation is performed to obtain the tth prediction result y (t) ; Repeat T times to obtain {y (1) ,y (2) ...y (T)}; Calculate the predicted mean y mean =sum(y (t) ) / T; Calculate the prediction variance y var =sum((y (t) -y mean ) 2 For example, for the electricity price forecast at a certain moment, the mean of 100 sampling results is 0.824 yuan / kWh, and the variance is 0.0009 (yuan / kWh). 2 , with a standard deviation of 0.03 yuan / kWh.

[0033] For the original input data X, first calculate the input gradient g=∂L / ∂X, that is, the gradient of the loss function L with respect to the input X; normalize the gradient g norm =g / ||g||, where ||g|| represents the Euclidean norm of the gradient; dynamically determine the perturbation strength ε according to the predicted standard deviation, specifically ε=α×std(y), where α is the perturbation coefficient (set to 2) and std(y) is the predicted standard deviation; generate adversarial samples X'=X+ε×g normDiscrete features (such as date type) remain unchanged and only continuous features are perturbed.

[0034] For example, if the original electricity price feature vector is [0.4, 0.4, 0.8, 1.2...] yuan / kWh, the prediction standard deviation is 0.03 yuan / kWh, the perturbation strength ε = 2 × 0.03 = 0.06 yuan / kWh, and the normalized gradient vector is [0.5, 0.5, -0.5, -0.5...], then the adversarial example electricity price features are [0.4 + 0.06 × 0.5, 0.4 + 0.06 × 0.5, 0.8 + 0.06 × (-0.5), 1.2 + 0.06 × (-0.5)...] = [0.43, 0.43, 0.77, 1.17...] yuan / kWh. During training, 5 adversarial examples are generated per batch and mixed with 20 original examples in a ratio of 1:4.

[0035] The confidence interval calculation process uses precise statistical methods based on the prediction result set {y (1) ,y (2) ...y (T)}, sort in ascending order to get the sorted sequence {y [1] ,y [2] ...y [T]}; For a 95% confidence interval, calculate the lower boundary index low =floor(0.025×T)=2, upper boundary index high =ceil(0.975×T)=98; the lower limit of the confidence interval is y [2] , the upper limit is y

[98] .

[0036] For example, after the 100 sample values ​​of electricity price forecast at a certain moment are arranged in ascending order, the second smallest value is 0.768 yuan / kWh and the 98th largest value is 0.879 yuan / kWh, then the 95% confidence interval is [0.768, 0.879] yuan / kWh. In addition, a parameterized method can also be used based on the predicted mean y mean = 0.824 yuan / kWh and the predicted standard deviation std(y) = 0.03 yuan / kWh, the 95% confidence interval is calculated as [y mean -1.96×std(y), y mean +1.96×std(y)]=[0.765,0.883] yuan / kWh. The intervals obtained by the two methods are close, verifying the consistency of the calculations.

[0037] The dynamic adjustment coefficient is calculated using a piecewise function, taking into account the nonlinear effect of uncertainty, and defining the lower threshold σ of the prediction standard deviation low =0.02, upper threshold σ high=0.1, maximum adjustment coefficient β max =0.2; when the prediction standard deviation std(y)≤σ low When std(y)≥σ high When β=β max ; When σ low <std(y)<σ high When β=β max ×(std(y)-σ low ) / (σ high -σ low ).

[0038] For example, when the forecast standard deviation is 0.05, the adjustment coefficient β = 0.2 × (0.05-0.02) / (0.1-0.02) = 0.075. The adjustment direction is determined by the historical forecast deviation statistics, specifically by calculating the average forecast deviation bias = average (y predict -y true If bias > 0, the forecast is generally biased upward, and the adjustment is negative; if bias < 0, the forecast is generally biased downward, and the adjustment is positive. Assuming bias = 0.04 > 0, the forecast mean is 0.824 yuan / kWh, and the adjustment factor is 0.075, the adjusted forecast value is 0.824 - 0.075 × 0.824 = 0.762 yuan / kWh.

[0039] The complete algorithm flow of the revenue calculation module includes multiple steps, and the input includes the optimized electricity price forecast p (t) , Power Forecast Plan P (t) and confidence interval [p low (t) ,p high (t) ]、[P low (t) ,P high (t) ], t=1,2...24 represents 24 hours. Positive power value indicates charging, negative power value indicates discharging. Assume that the charge and discharge efficiency η=0.9, the state of charge constraint SOC min =0.1, SOC max =0.9, the initial SOC value is SOC0=0.5, and the battery capacity is E=5MWh.

[0040] Verify that the power plan meets the SOC constraint. For t=1,2...24, calculate the SOC (t) =SOC (t-1) +P (t) ×Δt / E, where Δt=1 hour, if SOC (t) <SOCmin or SOC (t) >SOC max , then adjust P (t) Make SOC (t) Satisfy the constraints; calculate the charge and discharge power in each period, the charging power E charge (t) =max(0,P (t) )×Δt, discharge capacity E discharge (t) =max(0,-P (t) )×Δt×η; calculate the cost and benefit of each period, charging cost C (t) =E charge (t) ×p (t) , discharge gain R (t) =E discharge (t) ×p (t) / η; calculate the total cost C=sum(C (t) ), total revenue R=sum(R (t) ), Net Income R =RC; calculate the profit interval based on the confidence interval, the minimum net profit Net R low Using the electricity price floor and power floor to calculate the maximum net profit Net R high Calculated using electricity price caps and power caps.

[0041] For example, during the off-peak period (0:00-6:00), the predicted electricity price is 0.4 yuan / kWh, the planned charging power is 1MW, the charging capacity is 6MWh, and the cost is 2400 yuan. During the peak period (18:00-22:00), the predicted electricity price is 1.2 yuan / kWh, the planned discharge power is 1.2MW, and the discharge capacity is 4.8MWh. After considering efficiency, the actual discharge capacity is 5.33MWh, resulting in a profit of 6400 yuan. The net profit is 6400-2400 = 4000 yuan. Considering the electricity price confidence intervals of [0.38, 0.42] yuan / kWh and [1.15, 1.25] yuan / kWh, and the power confidence intervals of [0.9, 1.1]MW and [1.1, 1.3]MW, the calculated profit range is [3539, 4548] yuan.

[0042] In actual application at a 100MW / 200MWh large-scale commercial energy storage power station, this method was applied to real-time optimization decisions over 30 days. The operating strategy guided by the traditional deterministic model achieved an average daily profit of 285,000 yuan, while the operating strategy optimized by this method achieved an average daily profit of 342,000 yuan, an increase of 20.0%. At the same time, the uncertainty quantification provided by this method enables power station operators to adopt more conservative strategies for high-risk periods (such as when the forecast interval is wide), avoiding three potential major decision-making errors and recovering an estimated loss of more than 1.2 million yuan. This method has been operating stably at the power station for six months, increasing cumulative profits by 10.26 million yuan, demonstrating its practicality and effectiveness.

[0043] In an optional embodiment, a variational inference method is used to establish a Gaussian distribution model of the network parameters, and the prediction uncertainty of the multi-layer feedforward neural network structure is quantitatively characterized by the mean parameter and variance parameter of the Gaussian distribution model, including: Obtaining training data for a multi-layer feedforward neural network structure, constructing a likelihood function based on the training data, selecting a Gaussian distribution as a prior distribution of network parameters, and combining the likelihood function with the prior distribution using a Bayesian formula to construct a posterior distribution; Approximating the posterior distribution using a variational inference method to construct a variational distribution family in the form of a Gaussian distribution, wherein the variational distribution family includes a mean parameter vector and a variance parameter vector, and obtaining optimal variational distribution parameters by minimizing the Kullback-Leibler divergence between the variational distribution family and the posterior distribution using the variational inference method; Inputting the optimal variational distribution parameters into the multi-layer feedforward neural network structure, and for each layer in the multi-layer feedforward neural network structure, constructing the parameter distribution of the layer based on the mean parameter vector and the variance parameter vector using the variational inference method; obtaining network parameter samples based on the parameter distribution, and inputting the network parameter samples into the multi-layer feedforward neural network structure to obtain corresponding network outputs; Repeatedly obtain multiple groups of network parameter samples, input the multiple groups of network parameter samples into the multi-layer feedforward neural network structure to obtain multiple groups of network outputs, calculate the prediction mean and the prediction variance based on the multiple groups of network outputs; and use the prediction mean and the prediction variance as the prediction uncertainty quantification results of the multi-layer feedforward neural network structure.

[0044] Training data for multi-layer feedforward neural network structures is typically obtained from historical datasets for a specific task. For example, in a power load forecasting task, the training data consists of input features and target output values. Input features include historical load values, temperature, humidity, wind speed, date type, and other factors, while the target output is the load value at a future time. This dataset contains 10,000 training examples, each with 24 features and one target value. Data preprocessing includes normalization, mapping each feature value to a distribution with a mean of 0 and a standard deviation of 1 to minimize the impact of dimensionality differences on model training.

[0045] The likelihood function is constructed based on a preset error distribution model. For regression tasks, it is usually assumed that the error between the predicted value and the true value follows a Gaussian distribution with a mean of 0 and a constant variance. In the specific implementation, given the input feature x and the corresponding true output y, the neural network outputs the predicted value f(x), then the likelihood function is expressed as y follows a Gaussian distribution with f(x) as the mean and σ 2 is a Gaussian distribution with a variance of . For the entire training set, the likelihood function is the product of the likelihoods of each sample. In actual calculations, the logarithm is often converted to a summation form to improve numerical stability.

[0046] The prior distribution for network parameters uses a Gaussian distribution, specifically a standard Gaussian distribution with mean 0 and variance 1. This was chosen for its favorable mathematical properties, including analytical convenience and computational efficiency when combined with the likelihood function. Within the Bayesian framework, the posterior distribution is calculated using the Bayesian formula, expressed as the product of the prior distribution and the likelihood function, divided by the marginal likelihood. However, precisely calculating the posterior distribution is often difficult in complex neural networks, so variational inference methods are introduced for approximation.

[0047] The core idea of ​​the variational inference method is to find a parameterized distribution to approximate the true posterior distribution and construct a variational distribution family in the form of a Gaussian distribution. For each weight parameter w in the neural network, its variational distribution is set to have a mean of μ and a variance of σ. 2 The variational distribution family consists of a mean parameter vector μ and a variance parameter vector σ 2 Common description. Take a neural network with two hidden layers as an example. If the first layer has 100 neurons, the second layer has 50 neurons, the output layer has 1 neuron, and the input feature dimension is 24, then the total number of parameters is (24×100)+(100×50)+(50×1)=7450. The corresponding variational distribution family requires 7450 mean parameters and 7450 variance parameters.

[0048] Minimizing the KL divergence between the family of variational distributions and the posterior distribution is achieved using stochastic gradient descent. The actual optimization objective is to maximize the evidence lower bound (ELBO), which is defined as the log marginal likelihood minus the KL divergence. The optimization process uses a reparameterization technique, representing parameter sampling as the product of the mean parameter plus the variance parameter and samples from a standard normal distribution, allowing gradients to propagate through the random variables. Optimization uses the Adam optimizer with a learning rate of 0.001, a batch size of 128, and 10,000 training iterations. Through iterative optimization, the mean and variance parameter vectors gradually converge to optimal values. For example, the partial mean parameters of the first layer weights converge to [-0.232, 0.451, -0.178, 0.321], and the variance parameters converge to [0.021, 0.034, 0.018, 0.025].

[0049] When applying variational inference to each layer of a multi-layer feedforward neural network structure, the hierarchical Bayesian model idea is adopted. For the connection from the input layer to the first hidden layer, based on the mean parameter μ1 and the variance parameter σ1 2 Construct the weight distribution; for the connection from the first hidden layer to the second hidden layer, use μ2 and σ2 2 Construct the weight distribution; and so on until the output layer. The bias term of each layer is also modeled using a Gaussian distribution. In practical applications, a diagonal Gaussian distribution can be used to simplify the calculation, assuming that the different weight parameters are independent of each other.

[0050] The network parameter samples are obtained by using the Monte Carlo sampling method, which randomly extracts samples from the variational distribution of each weight parameter to form a complete set of network parameters. 2 The Gaussian distribution of is used, and the parameter sample is calculated as w = μ + σ × ε, where ε is a random number drawn from a standard normal distribution. For example, if a weight parameter has a mean of 0.5 and a variance of 0.04, then after drawing a random number 0.8 from a standard normal distribution, a sample of that weight is 0.5 + (0.2 × 0.8) = 0.66.

[0051] Repeat the above sampling process T times to obtain T sets of different network parameter samples. In practice, T is usually set between 10-100, and in this embodiment it is set to 50. These parameter samples are input into the neural network respectively, and 50 forward propagations are calculated for the same input feature x to obtain 50 different output values ​​{y1, y2...y 50 Based on these 50 output values, the prediction mean and prediction variance are calculated. The prediction mean is the arithmetic average of these output values, and the prediction variance is the variance of these output values. For example, the 50 prediction results for a test sample are [245.3, 251.8, 247.2, ..., 249.5] MW, and the calculated prediction mean is 248.7 MW and the prediction variance is 12.6 MW.2 .

[0052] The prediction mean and prediction variance, as quantitative measures of prediction uncertainty, have clear practical significance. The prediction mean represents the model's estimate of the target variable, while the prediction variance characterizes the degree of prediction uncertainty. A larger prediction variance indicates a lower confidence level in the model's prediction. This uncertainty quantification can be converted into a prediction interval in practical applications. For example, a 95% confidence interval can be expressed as the prediction mean ± 1.96 × the prediction standard deviation.

[0053] Applying this method to power load forecasting has improved prediction accuracy by 12.3% compared to traditional deterministic neural networks, while also providing reliable uncertainty estimates. In areas of high uncertainty (such as holidays and extreme weather), where traditional models often exhibit large errors, this method, through its larger prediction variance, provides early warnings, helping decision makers adopt more conservative scheduling strategies. Experiments show that this method achieves a 94.2% coverage rate within the prediction interval, close to the theoretical value of 95%, demonstrating the effectiveness of uncertainty quantification.

[0054] In actual deployment, this method requires multiple forward propagations during the prediction phase, which is computationally expensive. To address this issue, a Monte Carlo dropout method can be used to simplify the implementation. Specifically, dropout is kept active during prediction, and each forward propagation is equivalent to sampling a single instance from the model distribution. This variant improves computational efficiency by approximately 40% while maintaining the effectiveness of uncertainty quantification. On edge computing devices, performance can be further optimized by reducing the number of samples and compressing the model, thereby reducing resource consumption while maintaining prediction quality.

[0055] In an optional embodiment, the health status data of the energy storage device is analyzed by the deep reinforcement learning model, and a multi-dimensional life loss coefficient matrix is ​​established based on the number of cycles, deep cycle rate, and operating temperature to achieve an assessment of the device life loss, including: Acquire health status data of the energy storage device, construct a state feature vector based on the health status data, and use the state feature vector as input data for a deep reinforcement learning method; Based on the state feature vector, the deep reinforcement learning method is used to extract the number of cycles, deep cycle rate, and operating temperature parameters of the energy storage device. Based on the deep reinforcement learning method, a time decay factor is introduced to adaptively weight the operating data at different times in the state feature vector to obtain weighted historical loss data; Analyzing the dynamic change characteristics of the number of cycles, the deep cycle rate, and the operating temperature parameters based on the weighted historical loss data and inputting the dynamic change characteristics into a loss benchmark model to obtain a benchmark loss characteristic of the energy storage device; Based on the benchmark loss characteristics, a multi-dimensional life loss coefficient matrix is ​​constructed using the deep reinforcement learning method, and the multi-dimensional life loss coefficient matrix is ​​optimized by minimizing the deviation between the predicted loss value and the actual loss value to obtain an optimized loss coefficient matrix; The optimized loss coefficient matrix is ​​integrated with time to obtain a cumulative loss state of the energy storage device, and the remaining life of the energy storage device is determined based on the cumulative loss state and the design life of the energy storage device.

[0056] During the acquisition of energy storage device health status data, multi-dimensional parameters such as voltage, current, temperature, state of charge (SOC), and depth of discharge (DOD) are collected during device operation. The collection frequency is once every 5 minutes, with a total of 288 data points collected daily. Device operating parameters are stored according to different time spans, including minute-level, hourly-level, daily-level, and monthly-level data. Minute-level data primarily includes real-time voltage, current, power, and temperature parameters. For example, the data collected at a certain moment is: voltage 724.5V, current 56.2A, ambient temperature 25.3°C, battery pack temperature 28.7°C, and SOC value 72.5%. Hourly-level data primarily includes information such as average depth of discharge within the hour, maximum temperature, minimum temperature, and number of charge and discharge cycles. Daily-level data provides comprehensive statistics such as the entire day's charge and discharge energy, equivalent cycle count, and maximum depth of discharge.

[0057] When constructing the state feature vector, the collected raw data is preprocessed, including missing value processing, outlier filtering, and normalization. For missing values, the mean interpolation method of the previous and next moments is used to fill them; for outliers, the 3 times standard deviation method is used for screening, and data points out of the range are replaced with the nearest valid value. The normalization process uses the maximum and minimum value normalization method to map all features to the range of 0-1. The normalized data constitutes the state feature vector, the specific structure of which includes: timestamp features (year, month, day, hour, minute, second), environmental features (ambient temperature, humidity), battery operation features (SOC, DOD, number of cycles, charge and discharge power), and historical cumulative features (cumulative number of cycles, cumulative number of full charge and discharge, cumulative number of shallow charge and shallow discharge). For example, the state feature vector at a certain moment is [0.25, 0.42, 0.83, 0.12, 0.65, 0.33, 0.78, 0.56, 0.45, 0.68, 0.23, 0.44, 0.35, 0.67, 0.89], which correspond to the normalized parameter values ​​respectively.

[0058] The deep reinforcement learning method uses a dual-channel neural network structure, consisting of a feature extraction network and a state evaluation network. The feature extraction network consists of three convolutional layers and two fully connected layers. The convolutional layers are used to capture temporal features in time series data. The convolution kernel sizes are 5×5, 3×3, and 3×3, respectively, and the number of feature maps is 32, 64, and 64, respectively. Each layer is followed by a max pooling layer and a Reluctant Unit (ReLU) activation function. The fully connected layers have 256 and 128 neurons, respectively, for comprehensive extraction of high-level feature representations. The state evaluation network consists of three fully connected layers with 128, 64, and 32 neurons, respectively. The output layer contains multiple estimates, including the number of cycles, the depth cycle rate, and the temperature impact factor.

[0059] The number of cycles, deep cycle rate, and operating temperature parameters are extracted using a sliding window method with a window size of 24 hours and a step size of 1 hour to extract parameter changes within the past 24 hours. The number of cycles is defined as a complete cycle in which the battery is discharged again after being charged from a discharged state. A cycle is recorded when the SOC curve has a complete peak and trough. The deep cycle rate is defined as the difference between the highest SOC and the lowest SOC in a cycle. For example, when discharging from SOC 90% to SOC 30%, the deep cycle rate is 60%. The operating temperature parameters include the average temperature, the maximum temperature, and the temperature fluctuation amplitude, which are calculated by obtaining the battery surface temperature through the sensor.

[0060] The time decay factor uses an exponential decay approach, with data closer to the current time receiving a greater weight and data further away receiving a smaller weight. Specifically, for data at the i-th time point before time t, its weight is calculated as the base weight multiplied by the i-th power of the decay coefficient. For example, assuming a base weight of 1 and a decay coefficient of 0.95, the weight of the first historical point is 0.95, the weight of the second historical point is 0.95^2 = 0.9025, and the weight of the tenth historical point is approximately 0.599. After weighting, the effective weight of data within the last 24 hours exceeds 80%, ensuring that the model focuses more on recent operational status.

[0061] When weighting historical loss data, differentiated treatment of different types of losses is considered. Linear weighting is used for losses caused by the number of cycles; quadratic weighting is used for losses caused by the depth of cycling rate; and exponential weighting is used for losses caused by temperature. For example, a storage device has 2.5 cycles in 24 hours, an average depth of cycling rate of 55%, and an average operating temperature of 32°C. The weight of the most recent cycle is 1, and the weight of the previous day is 0.9. The weighted equivalent number of cycles is 2.5 × 1 + 2.2 × 0.9 = 4.48, the weighted equivalent depth of cycling rate is 55% × 1 + 60% × 0.9 = 109%, and the weighted equivalent temperature is 32°C × 1 + 28°C × 0.9 = 57.2°C.

[0062] During the analysis of dynamic change characteristics, the sliding variance method is used to detect parameter volatility, the trend analysis method is used to detect parameter change trends, and the cycle analysis method is used to detect parameter cyclical change characteristics. The sliding variance method calculates the variance of each parameter over the past seven days. Large variance values ​​indicate large parameter fluctuations and poor stability. Trend analysis uses linear regression to calculate the slope of parameter changes over the past 30 days. A positive slope indicates an upward trend in the parameter, while a negative slope indicates a downward trend. Cycle analysis uses the fast Fourier transform method to identify the main cycles of parameter changes.

[0063] The loss benchmark model is built using a random forest regression method based on a large amount of pre-collected historical data. Input features include parameters such as cycle number, deep cycle rate, and operating temperature, and the output is the predicted capacity loss rate. The model uses 80% of the historical data for training and 20% for validation. For example, a certain lithium battery model loses approximately 0.8% capacity after 100 cycles under standard conditions (25°C, 80% DOD), and approximately 7.5% after 1000 cycles. The benchmark model uses these loss data under standard conditions, combined with temperature correction factors and deep cycle correction factors, to predict capacity loss under actual operating conditions.

[0064] A multidimensional lifetime loss coefficient matrix was constructed using a three-dimensional grid method. The first dimension was the number of cycles, divided into 10 intervals (0-100, 100-300, 300-500, 500-1000, 1000-2000, 2000-3000, 3000-5000, 5000-8000, 8000-12000, and >12000). The second dimension was the depth of cycling, divided into five intervals (0-20%, 20-40%, 40-60%, 60-80%, and 80-100%). The third dimension was the operating temperature, divided into seven intervals (<0°C, 0-10°C, 10-20°C, 20-30°C, 30-40°C, 40-50°C, and >50°C). For each grid point, the loss coefficient under those conditions was predicted using deep reinforcement learning. For example, when the number of cycles is 1500, the deep cycle rate is 55%, and the operating temperature is 35°C, the loss coefficient is 0.00085, which means that the battery capacity loses about 0.085% for each cycle under these conditions.

[0065] The loss coefficient matrix is ​​optimized using gradient descent, with parameters adjusted based on the mean squared error between the predicted and actual loss values. The learning rate is set to 0.001, the batch size is 64, and the number of training iterations is 1000. The error between the predicted and actual loss is calculated at each iteration, and the parameters in the loss coefficient matrix are updated via backpropagation. For example, under certain conditions, the initial predicted loss is 0.75%, the actual loss is 0.82%, and the error is 0.07%. After optimization, the predicted loss is adjusted to 0.80%, reducing the error to 0.02%. The average prediction error of the optimized loss coefficient matrix on the test set is reduced from an initial 12.5% ​​to 4.3%.

[0066] During the cumulative loss state calculation process, the losses are integrated over the time series. For discrete time points, the trapezoidal integration method is used for approximate calculations. Specifically, the loss coefficient for each time period is multiplied by the number of cycles within that time period and then accumulated. For example, in the first month, the energy storage device cycled 45 times under various conditions, with an average loss coefficient of 0.00075. The capacity loss for the first month is 45 × 0.00075 = 0.03375, or 3.375%. The losses are calculated and accumulated for each month to obtain the total cumulative loss.

[0067] The remaining life is determined by comparing the cumulative loss state with the design life. Taking capacity decay as an example, if the battery's design life is 80% of the initial capacity (allowing for 20% capacity decay) and the current cumulative loss is 8%, the remaining life percentage is (20% - 8%) / 20% = 60%. If the battery's design cycle life is 6,000 cycles, the estimated remaining number of cycles is approximately 6,000 × 60% = 3,600 cycles. The remaining life can be further estimated based on the current frequency of use. For example, if the current average daily cycle count is 2.5, the estimated remaining life is approximately 3,600 / 2.5 / 365 = 3.95 years.

[0068] In practice, this method was applied to a grid-side energy storage power station, comprising a 10MWh lithium-ion battery energy storage system. After 1.5 years of operation, the system's cumulative capacity loss, as assessed by this method, was 5.8%, resulting in a predicted remaining lifespan of 14.2 years. This agreement exceeded 95% with the actual capacity test result (5.5% loss), demonstrating the effectiveness of the method. Compared to traditional lifespan assessment methods, this method achieved 25% higher prediction accuracy, particularly under complex operating conditions.

[0069] In an optional embodiment, the deep reinforcement learning method is used to construct a multidimensional life loss coefficient matrix, and the multidimensional life loss coefficient matrix is ​​optimized by minimizing the deviation between the predicted loss value and the actual loss value. The optimized loss coefficient matrix includes: A multi-dimensional life loss coefficient matrix includes a cycle number dimension, a deep cycle rate dimension, and a temperature dimension; minimizing the deviation between the predicted loss value output by the multi-dimensional life loss coefficient matrix and the actual loss value; The multi-dimensional life loss coefficient matrix is ​​optimized based on the result of the minimization process to obtain an optimized loss coefficient matrix.

[0070] The process of constructing a multi-dimensional lifetime loss coefficient matrix begins with data collection and preprocessing. Battery loss data is collected under different cycle times, cycle depths, and temperature conditions. For example, at room temperature of 25°C and a cycle depth of 80%, the battery's capacity loss after 100, 200, and 500 cycles is 5%, 8%, and 15%, respectively. At a high temperature of 40°C and a cycle depth of 60%, the battery's capacity loss after the same number of cycles is 7%, 11%, and 20%, respectively. This data is standardized and used as the training dataset for deep reinforcement learning.

[0071] The deep reinforcement learning network architecture consists of three core components: state space, action space, and reward function. The state space is defined as the current state of the battery, including the number of cycles it has experienced, the deep cycle rate, and the ambient temperature. For example, the state can be represented as the battery having undergone 300 cycles, with an average deep cycle rate of 70%, and an average ambient temperature of 30°C. The action space is defined as the adjustment operation on the multi-dimensional life loss coefficient matrix, including increasing or decreasing the coefficient value on a specific dimension. The reward function is designed to be the negative value of the deviation between the predicted loss value and the actual loss value. The smaller the deviation, the greater the reward.

[0072] During training, a deep Q-network (DQN) algorithm was used to implement deep reinforcement learning. The network input was the battery state information, and the output was the adjustment action for the loss coefficient matrix. When initializing the multi-dimensional lifetime loss coefficient matrix, the cycle number dimension was set to [0.05, 0.08, 0.12, 0.18, 0.25], corresponding to 100, 200, 500, 1000, and 2000 cycles, respectively; the deep cycling rate dimension was set to [0.6, 0.8, 1.0, 1.2, 1.4], corresponding to deep cycling rates of 20%, 40%, 60%, 80%, and 100%, respectively; and the temperature dimension was set to [0.7, 0.9, 1.0, 1.2, 1.5], corresponding to ambient temperatures of 0°C, 15°C, 25°C, 35°C, and 45°C, respectively.

[0073] The Deep Q Network uses a four-layer neural network structure, consisting of an input layer, two hidden layers, and an output layer. The number of nodes in the input layer corresponds to the state space dimension, while the number of nodes in the output layer corresponds to the action space dimension. The number of nodes in the two hidden layers is 128 and 64, respectively. The ReLU function is used as the activation function, the Adam algorithm is used as the optimizer, and the learning rate is set to 0.001.

[0074] The training process uses an experience replay mechanism. The experience pool size is set to 10,000, and 128 samples are randomly sampled from the experience pool for batch training each time. The discount factor is set to 0.95, the initial ε value of the ε-greedy strategy is 1.0, the minimum ε value is 0.01, and the ε decay rate is 0.995. The number of training iterations is set to 5,000, and the model performance is evaluated every 100 iterations.

[0075] During training, the deep reinforcement learning algorithm continuously optimizes the coefficient values ​​of each dimension in the matrix by trying different loss coefficient matrix adjustment actions. For example, if it detects that the predicted loss value is significantly lower than the actual loss value in a high temperature environment (such as 40°C), the algorithm will increase the coefficient value of the temperature dimension in the high temperature area; if it finds that the predicted loss value is higher than the actual loss value at a large number of cycles (such as 2000), the algorithm will reduce the coefficient value of the number of cycles dimension in the large number of cycles area.

[0076] The algorithm optimizes the multidimensional lifetime loss coefficient matrix by minimizing the deviation between predicted and actual loss values, using the mean squared error (MSE) as the loss function. At each iteration, the predicted loss value is calculated based on the current loss coefficient matrix and then compared with the actual loss value to calculate the mean squared error. Based on the magnitude and direction of the error, the algorithm adjusts the parameters of the deep Q network through backpropagation, which in turn guides the adjustment of the loss coefficient matrix.

[0077] After 5000 training iterations, an optimized multi-dimensional lifetime loss coefficient matrix was obtained. The optimized cycle number dimension coefficients became [0.048, 0.079, 0.125, 0.176, 0.242], the depth cycle rate dimension coefficients became [0.58, 0.82, 1.03, 1.18, 1.45], and the temperature dimension coefficients became [0.68, 0.88, 1.0, 1.25, 1.58]. The optimized loss coefficient matrix achieved a prediction accuracy of 92.5% on the test dataset, with a mean absolute error reduced to 2.3%, a significant improvement over the initial matrix's 85.0% accuracy and 4.7% mean absolute error.

[0078] To verify the effectiveness of the optimized loss coefficient matrix, the predicted and actual loss values ​​under different conditions were compared. For example, at room temperature of 25°C, a deep cycling rate of 80%, and 800 cycles, the predicted loss value before optimization was 16.2%, while the actual loss value was 14.8%, with an absolute error of 1.4%. After optimization, the predicted loss value was 15.0%, with the absolute error reduced to 0.2%. At high temperature of 40°C, a deep cycling rate of 100%, and 1500 cycles, the predicted loss value before optimization was 28.5%, while the actual loss value was 32.0%, with an absolute error of 3.5%. After optimization, the predicted loss value was 31.2%, with the absolute error reduced to 0.8%.

[0079] The multi-dimensional life loss coefficient matrix constructed and optimized through deep reinforcement learning methods can accurately predict the battery life loss under different operating conditions, providing effective support for battery life prediction and health management.

[0080] In an optional embodiment, multiple groups of candidate charge and discharge scheduling strategies are generated based on the deep reinforcement learning model, with predicted revenue and life loss as dual constraints of the reward function. By alternately training the policy network and the value network, the generation of charge and discharge strategies that maximize revenue while considering the life of the equipment is achieved, including: A strategy generation network is constructed based on a deep reinforcement learning model. The strategy generation network includes a parameter generation module and a value assessment module. The network uses the grid load status, electricity price information, device charge status, and ambient temperature of the energy storage device as the system state to generate corresponding charge and discharge action strategy parameters. The system state and the charge and discharge action strategy parameters are combined to form a state-action pair. constructing a dual-constraint reward function based on the state-action pair, inputting the system state into the parameter generation module to obtain a mean parameter and a standard deviation parameter of the action distribution, and inputting the system state and the dual-constraint reward function into the value evaluation module to obtain a state value; An advantage value is calculated based on the state value and the dual-constraint reward function, the advantage value is used to guide the parameter update of the policy generation network, and the temporal difference error of the state value is used to guide the parameter update of the policy generation network, and an optimized policy generation network is obtained through alternating training; The prediction result of the optimized strategy generation network is substituted into the dual-constraint reward function to obtain a strategy evaluation index. Based on the strategy evaluation index, the optimal charge and discharge strategy is selected under the condition of satisfying the life loss constraint to obtain the optimal charge and discharge control strategy of the energy storage device.

[0081] The strategy generation network uses a parameterized Gaussian distribution to represent strategies. It consists of two core components: a parameter generation module and a value assessment module. The parameter generation module consists of a three-layer fully connected neural network. The input layer has the same number of neurons as the state dimension (including grid load status, electricity price information, device state of charge, and ambient temperature), the hidden layer has 256 neurons, and the output layer has twice the number of neurons as the action dimension, representing the mean and standard deviation parameters of the action distribution, respectively. The input state consists of 48-dimensional electricity price forecast data (2 data points per hour for 24 hours), 24-dimensional load forecast data, 1-dimensional current SOC value, and 1-dimensional ambient temperature, for a total of 74 dimensions. The output is 48-dimensional action parameters, corresponding to the charging and discharging power decisions for each hour within the 24-hour period.

[0082] The value assessment module also uses a three-layer fully connected neural network. The input layer is the same as the parameter generation module, with 256 hidden layer neurons and a single output neuron, representing the estimated value of the current state. The ReLU activation function is used to enhance the network's expressiveness, while the final layer does not use an activation function to ensure an unrestricted output range. Network parameters are updated using stochastic gradient descent, with a learning rate of 0.0003 and a batch size of 64.

[0083] The grid load status, electricity price information, device state of charge, and ambient temperature are combined into a state vector and input into the parameter generation module to obtain the action distribution parameters. The state of a 100kWh energy storage device includes information such as the current peak, valley, and average time period, an electricity price of 0.8 yuan / kWh, a SOC of 45%, and an ambient temperature of 28°C. The parameter generation module outputs an action distribution with a mean parameter of 35kW (positive values ​​indicate charging, negative values ​​indicate discharging) and a standard deviation of 5kW. Based on these parameters, a specific charge and discharge power decision, such as a charging power of 37kW, is generated through random sampling to form a state-action pair.

[0084] The dual-constrained reward function includes a revenue term and a lifetime loss term. The revenue term is calculated as the economic benefits generated by the charging and discharging process, while the lifetime loss term is calculated as the device capacity degradation caused by the charging and discharging process. The revenue term is the discharge revenue minus the charging cost. The discharge revenue is equal to the discharged energy multiplied by the electricity price at the time of discharge, while the charging cost is equal to the charged energy multiplied by the electricity price at the time of charge. The lifetime loss term constructs a loss model based on the cycle depth, temperature, and current rate. The greater the cycle depth, the higher or lower the temperature, and the higher the current rate, the more severe the lifetime loss.

[0085] A certain energy storage device charges 30 kWh at an electricity price of 0.3 yuan / kWh and discharges 27 kWh at an electricity price of 0.9 yuan / kWh (assuming 90% charge and discharge efficiency). The revenue is 27 × 0.9 - 30 × 0.3 = 24.3 - 9 = 15.3 yuan. This charge and discharge behavior results in a 0.015% capacity loss, which, based on the initial device cost of 150,000 yuan, is equivalent to a loss of approximately 22.5 yuan. After setting the weight coefficient, the final reward is the difference between the revenue and the weighted loss.

[0086] The weight ratio between the revenue and lifespan loss terms in the dual-constraint reward function is set as an adjustable parameter, typically initially set at 7:3. Early in the life of a device, the revenue weight can be increased to 8:2, while near the end of its lifespan, the lifespan loss weight can be increased to 4:6, achieving a dynamic balance. Weight adjustments can be dynamically set based on operational objectives, such as temporarily increasing the revenue weight to 9:1 when participating in demand response or temporarily increasing the lifespan weight to 3:7 after equipment maintenance.

[0087] The current state is input into the value assessment module to obtain an estimated value for that state. The reward value and next state are also obtained based on the action performed. The next state is input into the value assessment module to obtain an estimated value for the next state. The time series difference error (TSDE) is calculated based on the current reward value, current state value, and next state value. If the current state value is estimated to be 50, a charging action is performed to obtain a reward of 15, the next state value is estimated to be 60, and the discount factor is 0.95, then the TSDE is 15 + 0.95 × 60 - 50 = 22.

[0088] The advantage value is calculated by accumulating the discounted rewards for the next n steps and subtracting the estimated value of the current state. If n=5, the discounted rewards for the next 5 steps will be accumulated. The current reward is 15, and the rewards for the next 4 steps are 12, 10, 8, and 6 respectively. The discount factor is 0.95, so the accumulated discounted reward is 15+0.95×12+0.95 2 ×10+0.95 3 ×8+0.95 4 ×6=47.73. If the current state value is estimated to be 40, the advantage value is 47.73-40=7.73.

[0089] Based on the calculated advantage value, a policy gradient objective function is constructed, and the parameters of the parameter generation module are updated using stochastic gradient ascent. Each training batch contains 64 state-action pairs. The product of the log probability and the advantage value is calculated for each sample, and the average of these 64 products is used as the gradient update direction. The learning rate is set to 0.0003, and the Adam optimizer is used for parameter updates.

[0090] Based on the calculated temporal difference error, a mean squared error loss function is constructed, and stochastic gradient descent is used to update the parameters of the value assessment module. The square of the temporal difference error for each sample is calculated and averaged across all samples in the batch to form the loss function value. The learning rate is also set to 0.0003, and the Adam optimizer is used for parameter updates.

[0091] During each training cycle, the policy network parameters are first updated, followed by the value network parameters, and this process is repeated until convergence. The maximum training cycle is set to 10,000, and policy performance is evaluated every 500 cycles. Convergence is considered achieved when performance improvement is less than 1% over five consecutive evaluations. The training data is based on historical operating data from real energy storage power plants, including charging and discharging records from different seasons and load conditions throughout the year, totaling approximately 90,000 data points. 80% of this data is used for training and 20% for validation.

[0092] The optimized policy generation network is applied to the predicted state for the next 24 hours to determine the charge and discharge power decision for that 24-hour period. Based on this decision, the revenue and lifespan loss are simulated and calculated. These revenue and lifespan loss are then substituted into the dual-constrained reward function to obtain the policy evaluation metric. For a policy predicting a total revenue of 180 yuan and a lifespan loss cost of 60 yuan within 24 hours, with a weighting ratio of 7:3, the evaluation metric is 0.7 × 180 - 0.3 × 60 = 126 - 18 = 108.

[0093] To determine life loss constraints, a maximum daily allowable life loss threshold is set, typically 1 / 3650 of the device's total lifespan (assuming a 10-year device lifespan). The predicted life loss is then compared to the threshold. Given a device's design lifespan of 6000 cycles and a maximum daily allowable life loss of 1.65 equivalent cycles, a strategy predicting a life loss of 1.8 equivalent cycles would not meet the constraint and would be excluded.

[0094] All candidate strategies that meet the lifespan constraints are sorted from highest to lowest based on their evaluation metrics, and the strategy with the highest metric is selected as the final execution strategy. Typically, 10-20 candidate strategies are generated, and after screening, 5-10 strategies that meet the constraints are retained. From these, the strategy with the highest evaluation metric is selected for execution. The execution strategy includes hourly charge and discharge power instructions for 24 hours, such as 30kW charging power from 0-6 am, 45kW discharging power from 10-14 am, and 40kW discharging power from 18-22 am.

[0095] The practical application of this method at a 10MWh / 5MW commercial energy storage power station demonstrated significant improvements in economic returns by 18.7% and equipment life expectancy by 23.5%, compared to traditional single-objective optimization methods. The advantages of this method were particularly pronounced during the summer months, when electricity prices fluctuate significantly and temperatures fluctuate dramatically. Daily returns increased by over 25%, while effectively controlling equipment lifespan losses.

[0096] Figure 3 This is a trend diagram of the convergence and reward function changes of the deep reinforcement learning model of an embodiment of the present invention. The figure shows the changing trends of four key indicators during the training process: the profit reward (circular mark) reflecting the economic benefits of charging and discharging of the energy storage equipment, the life loss penalty (square mark) representing the impact of the charging and discharging strategy on the equipment life, the total reward value (triangle mark) that comprehensively considers economic benefits and life loss, and the model loss value (dashed line) indicating the model training status. The entire training process is divided into three phases: In the exploration phase (0-1000 runs), the reward increases rapidly from 0 to approximately 150, the lifespan penalty gradually improves from -75 to -40, the total reward increases from -50 to 100, and the model loss continuously decreases from 0.2 to 0.8. In the convergence phase (1000-2500 runs), the growth rate of various indicators slows down, with the reward reaching approximately 230, the lifespan penalty approaching 0, the total reward rising to 200, and the model loss further decreasing to 1.2. In the stabilization phase (2500-5000 runs), all indicators stabilize, with the reward ultimately stabilizing at 270, the lifespan penalty at 20, the total reward remaining at 250, and the model loss decreasing to 1.8 and remaining stable. This training process fully demonstrates the convergence process of deep reinforcement learning models in optimizing energy storage device charging and discharging strategies by balancing economic benefits and lifespan loss to achieve continuous performance improvement.

[0097] In an optional embodiment, substituting the prediction result of the optimized strategy generation network into the dual-constraint reward function to obtain a strategy evaluation index, and selecting the optimal charge and discharge strategy based on the strategy evaluation index while satisfying the life loss constraint includes: Substituting the prediction results of the optimized strategy generation network into a dual-constraint reward function, the dual-constraint reward function including a revenue index and a life loss index of the energy storage device, and evaluating the prediction results based on the dual-constraint reward function to obtain a strategy evaluation index corresponding to the prediction results; Calculating a cumulative life loss value of the energy storage device based on the life loss index in the prediction result, and comparing the cumulative life loss value with a preset loss threshold to obtain a life loss constraint condition of the prediction result; The prediction results that meet the life loss constraint condition are sorted in descending order according to the size of the strategy evaluation index, and the results of the descending sorting are used as candidate charging and discharging strategies for the energy storage device; and the charging and discharging strategy with the largest strategy evaluation index among the candidate charging and discharging strategies is used as the optimal charging and discharging control strategy for the energy storage device.

[0098] The optimized policy generation network's prediction results are substituted into the dual-constraint reward function to calculate the policy evaluation metric. The dual-constraint reward function includes two core evaluation dimensions: revenue and lifespan loss. The revenue metric is calculated by calculating the revenue for each time period (e.g., 15 minutes) of charging and discharging within a complete scheduling cycle (typically 24 hours). For charging, the calculation is the charge amount multiplied by the current time period's electricity price, then multiplied by -1; for discharging, the calculation is the discharge amount multiplied by the current time period's electricity price. The total revenue metric is obtained by summing the revenue across all time periods and subtracting the fixed operation and maintenance costs.

[0099] Assuming a 100kWh / 50kW energy storage system charges 40kWh at 2:00 AM when the electricity price is 0.3 yuan / kWh, and discharges 36kWh at 2:00 PM when the electricity price is 0.9 yuan / kWh (assuming 90% charge and discharge efficiency), the profit from this operation is: 36kWh × 0.9 yuan / kWh - 40kWh × 0.3 yuan / kWh = 32.4 yuan - 12 yuan = 20.4 yuan. If the daily fixed operation and maintenance cost is 5 yuan, the net profit is 15.4 yuan.

[0100] Determine the battery's current state of charge (SOC), calculate the charge and discharge depth, query the life loss coefficient matrix, and calculate the life loss value for a single operation. The charge and discharge depth is the absolute value of the change in SOC before and after charge and discharge. For example, charging from 30% to 90% is a depth of 60%. Query the loss coefficient at the corresponding depth based on a pre-established life loss coefficient matrix. This coefficient matrix is ​​based on experimental data and records the capacity loss rate corresponding to different depth cycles. For example, the loss coefficient at a depth of 20% is 0.0025%, at a depth of 60% is 0.012%, and at a depth of 100% is 0.025%.

[0101] The life loss indicator also needs to consider temperature. The impact of temperature on loss is reflected in the temperature correction factor. This factor is 1.0 at 25°C. For every 10°C increase in temperature, the factor increases by 0.2; for every 10°C decrease in temperature, the factor increases by 0.15. For example, operating at 35°C, the temperature correction factor is 1.2; at 15°C, the factor is 1.15. The final life loss value is the basic loss factor multiplied by the temperature correction factor, then multiplied by the number of cycles.

[0102] Set the revenue weighting coefficient α and the life loss weighting coefficient β, typically α + β = 1, for example α = 0.7 and β = 0.3. Convert the revenue metric to a standardized revenue score by dividing the current revenue by the historical highest revenue and multiplying by 100, with a score range of 0-100. Similarly, convert the life loss metric to a standardized life score by subtracting the ratio of the current loss value to the maximum acceptable loss value from 1 and multiplying by 100, also with a score range of 0-100. The final reward function value is: α × standardized revenue score - β × (100 - standardized life score).

[0103] A strategy predicts a return of 200 yuan, and its historical maximum return is 250 yuan. Therefore, its standardized return score is (200 / 250) × 100 = 80 points. The strategy predicts a lifespan loss of 0.015%, and its maximum acceptable loss is 0.025%. Therefore, its standardized lifespan score is [1 - (0.015 / 0.025)] × 100 = 40 points. Using a weighting of α = 0.7 and β = 0.3, the final reward function value is 0.7 × 80 - 0.3 × (100 - 40) = 56 - 18 = 38 points.

[0104] Obtain the energy storage device's historical cumulative loss value, calculate the new loss value predicted by the current policy, and add the two together to obtain the updated cumulative loss value. This historical cumulative loss value is stored in the device management database and updated after each policy execution. For example, if an energy storage device has been operating for two years and its current cumulative loss value is 15%, and the new policy predicts a loss of 0.02%, the updated cumulative loss value is 15.02%.

[0105] Determine the total loss cap based on the equipment's design lifespan, determine phased loss limits based on operational planning, and establish a dynamic adjustment mechanism based on seasonality and grid demand. For example, for a lithium battery energy storage system with a designed capacity degradation cap of 20% and an expected service life of 10 years, the annual loss threshold is 2%, the monthly loss threshold is approximately 0.167%, and the daily loss threshold is approximately 0.0055%. To account for seasonal fluctuations in grid demand, the daily threshold can be adjusted to 0.007% during peak summer months and 0.004% during low winter months.

[0106] Calculate the estimated cumulative loss after the current policy is executed, compare this value with the preset threshold for the corresponding time period, and generate the constraint judgment result. For example, if the estimated cumulative loss after the policy is executed on a certain day is 15.02%, and that day is the 730th day of operation, the corresponding threshold is 730 × 0.0055% = 4.015%, and the initial capacity is 100%, then the actual current loss should be above 96%. The actual loss is 100% - 15.02% = 84.98%, which is significantly lower than the threshold requirement, and the policy does not meet the constraint condition.

[0107] After comparing the accumulated life loss value with a preset threshold, if it does not exceed the preset threshold, the strategy is retained as a valid candidate; if it exceeds the preset threshold, it is removed from the candidate set. In practice, a three-level threshold system can be set: warning threshold, limit threshold, and prohibition threshold. When the accumulated loss reaches the warning threshold (e.g., 70% of the design loss), an alert is issued; when it reaches the limit threshold (e.g., 85% of the design loss), only the low-loss strategy is allowed to be executed; when it reaches the prohibition threshold (e.g., 95% of the design loss), only the protective strategy is allowed to be executed.

[0108] Sort the evaluation indicators of the strategies that meet the lifespan constraint in descending order, construct a set of candidate strategies, calculate the comprehensive evaluation indicator value of each strategy, sort them from high to low by indicator value, and generate a list of sorted results. For example, suppose there are five candidate strategies after screening, with comprehensive evaluation indicators of 78, 65, 82, 59, and 71, respectively. The sorted results are 82, 78, 71, 65, and 59, and the corresponding strategies are numbered 3, 1, 5, 2, and 4.

[0109] The strategy with the highest evaluation metric is selected from the candidate strategy list in descending order as the final execution strategy. For example, in the above example, Strategy 3 has an evaluation metric of 82, ranking first and therefore being selected as the optimal charge and discharge control strategy. This strategy includes charge and discharge power instructions for each time period within a 24-hour period, such as 35kW charging power from 1:00 AM to 3:00 AM, 20kW discharging power from 10:00 AM to 12:00 AM, and 40kW discharging power from 2:00 PM to 4:00 PM.

[0110] The selected optimal strategy is converted into a standard control instruction format and sent via a communication interface to the energy storage plant's energy management system. The energy management system executes the instructions to control the charging and discharging behavior of each energy storage unit. For example, the selected strategy is converted into a JSON-formatted control instruction containing information such as timestamp, power value, and operation type. The instruction is sent to the energy management system via industrial Ethernet. The energy management system parses the instruction and controls the energy storage converter via the fieldbus to perform the corresponding charging and discharging operations.

[0111] A real-time monitoring mechanism is established to collect data on policy execution performance, analyze the deviation between policy and actual returns, and dynamically adjust the reward function weight parameters. The real-time monitoring mechanism collects data such as battery SOC changes, actual charge and discharge power, real-time grid electricity prices, and ambient temperature through the energy storage management system. Revenue and loss performance data are summarized daily, and the deviation rate between actual and predicted returns, as well as the deviation rate between actual and predicted lifetime losses, is calculated. When the deviation rate exceeds a preset threshold (e.g., ±10%), the model's automatic optimization mechanism is triggered, retraining the deep reinforcement learning model.

[0112] The dynamic weight adjustment mechanism automatically adjusts the weight coefficients in the dual-constraint reward function based on long-term operational data analysis. For example, in the early stages of equipment operation (e.g., the first three years), the emphasis is on profitability, with α set to 0.8 and β to 0.2. In the mid-term (e.g., 3-7 years), α is adjusted to 0.6 and β to 0.4. In the late stages (e.g., 7 years and above), α is adjusted to 0.4 and β to 0.6, placing greater emphasis on protecting the equipment's lifespan.

[0113] Define emergency types, develop emergency response strategies, and set up safety recovery procedures. These include sudden grid failures, abnormal battery temperatures, and battery SOC exceeding limits. For example, if the battery temperature exceeds 45°C, the charge and discharge power is immediately reduced to 50% of the rated value. If the temperature exceeds 55°C, charging and discharging are immediately stopped and forced cooling is initiated. If the temperature exceeds 65°C, the battery pack is immediately disconnected and the firefighting system is activated.

[0114] The above steps form a complete dual-constraint process for selecting the optimal charging and discharging strategy, ensuring that commercial energy storage power plants maximize economic benefits while protecting equipment life. This approach has been validated in multiple real-world energy storage projects, achieving an average 15% improvement in economic benefits and approximately 22% longer equipment life compared to traditional single-objective optimization methods.

[0115] A second aspect of an embodiment of the present invention provides an intelligent optimization system for a commercial energy storage power station revenue model, including: The first unit is configured to obtain historical operating data of the commercial energy storage power station, wherein the historical operating data includes charge and discharge power data, battery state of charge data, real-time grid electricity price data, and energy storage device health status data; A second unit is configured to construct a deep reinforcement learning model for a commercial energy storage power station based on the historical operating data, wherein the deep reinforcement learning model adopts a deep neural network structure and nonlinearly maps the historical operating data through the deep neural network structure to generate a predicted profit value; A third unit is configured to analyze the health status data of the energy storage device using the deep reinforcement learning model, establish a multi-dimensional life loss coefficient matrix based on the number of cycles, deep cycle rate, and operating temperature, and implement an assessment of the life loss of the device; A fourth unit is configured to generate multiple sets of candidate charge and discharge scheduling strategies based on the deep reinforcement learning model, using predicted revenue and life loss as dual constraints of the reward function, and to generate charge and discharge strategies that maximize revenue while taking into account equipment life through alternating training of the policy network and the value network; The fifth unit is used to generate an operation control instruction according to the optimized charge and discharge scheduling strategy, and send the operation control instruction to the energy management system of the commercial energy storage power station to control the energy storage power station to operate according to the optimized charge and discharge scheduling strategy.

[0116] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0117] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0118] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent optimization method for the revenue model of a commercial energy storage power station, characterized in that: include: Acquire historical operating data of the commercial energy storage power station, the historical operating data including charge and discharge power data, battery state of charge data, real-time grid electricity price data, and energy storage equipment health status data; Building a deep reinforcement learning model for a commercial energy storage power station based on the historical operating data, wherein the deep reinforcement learning model uses a deep neural network structure to nonlinearly map the historical operating data through the deep neural network structure to generate a predicted profit value; Analyzing the health status data of the energy storage device through the deep reinforcement learning model, establishing a multi-dimensional life loss coefficient matrix based on the number of cycles, deep cycle rate, and operating temperature, and realizing the assessment of the equipment life loss; Based on the deep reinforcement learning model, multiple groups of candidate charge and discharge scheduling strategies are generated. The predicted revenue and life loss are used as dual constraints of the reward function. By alternately training the policy network and the value network, the generation of charge and discharge strategies that maximize revenue while considering the life of the equipment is achieved. An operation control instruction is generated according to the optimized charge and discharge scheduling strategy, and the operation control instruction is sent to the energy management system of the commercial energy storage power station to control the energy storage power station to operate according to the optimized charge and discharge scheduling strategy.

2. The method according to claim 1, characterized in that A deep reinforcement learning model for a commercial energy storage power station is constructed based on the historical operating data. The deep reinforcement learning model adopts a deep neural network structure. The historical operating data is nonlinearly mapped by the deep neural network structure to generate a predicted profit value. The generated predicted profit value includes: The deep reinforcement learning model adopts a multi-layer feedforward neural network structure, which includes an input layer, a hidden layer, and an output layer, wherein the hidden layer performs feature extraction and nonlinear mapping on the historical operation data through a weight matrix, a bias vector, and a nonlinear activation function; The network parameters of the multi-layer feedforward neural network structure are constructed as random variables, a Gaussian distribution model of the network parameters is established using a variational inference method, and the prediction uncertainty of the multi-layer feedforward neural network structure is quantitatively characterized by the mean parameter and variance parameter of the Gaussian distribution model; Adding disturbances to the historical operating data based on the Gaussian distribution model to generate adversarial samples, and using the adversarial samples in the training process of the multi-layer feedforward neural network structure to improve the robustness of the multi-layer feedforward neural network structure; Calculating a prediction mean and a prediction standard deviation based on the prediction results of the multi-layer feedforward neural network structure, and determining a confidence interval of the prediction results according to the prediction mean and the prediction standard deviation; calculating a dynamic adjustment coefficient according to the prediction standard deviation, and adaptively optimizing the prediction results of the multi-layer feedforward neural network structure based on the dynamic adjustment coefficient to obtain an optimized prediction result; The optimized prediction result and the confidence interval are input into a profit calculation module, and a predicted profit value of the commercial energy storage power station is generated by the profit calculation module.

3. The method according to claim 2, characterized in that A Gaussian distribution model of the network parameters is established using a variational inference method, and the prediction uncertainty of the multi-layer feedforward neural network structure is quantitatively characterized by the mean parameter and variance parameter of the Gaussian distribution model. The method includes: Obtaining training data for a multi-layer feedforward neural network structure, constructing a likelihood function based on the training data, selecting a Gaussian distribution as a prior distribution of network parameters, and combining the likelihood function with the prior distribution using a Bayesian formula to construct a posterior distribution; Approximating the posterior distribution using a variational inference method to construct a variational distribution family in the form of a Gaussian distribution, wherein the variational distribution family includes a mean parameter vector and a variance parameter vector, and obtaining optimal variational distribution parameters by minimizing the Kullback-Leibler divergence between the variational distribution family and the posterior distribution using the variational inference method; Inputting the optimal variational distribution parameters into the multi-layer feedforward neural network structure, and for each layer in the multi-layer feedforward neural network structure, constructing the parameter distribution of the layer based on the mean parameter vector and the variance parameter vector using the variational inference method; obtaining network parameter samples based on the parameter distribution, and inputting the network parameter samples into the multi-layer feedforward neural network structure to obtain corresponding network outputs; Repeatedly obtain multiple groups of network parameter samples, input the multiple groups of network parameter samples into the multi-layer feedforward neural network structure to obtain multiple groups of network outputs, calculate the prediction mean and the prediction variance based on the multiple groups of network outputs; and use the prediction mean and the prediction variance as the prediction uncertainty quantification results of the multi-layer feedforward neural network structure.

4. The method according to claim 1, wherein The deep reinforcement learning model is used to analyze the health status data of the energy storage device, and a multi-dimensional life loss coefficient matrix is ​​established based on the number of cycles, deep cycle rate, and operating temperature to achieve an assessment of the device life loss, including: Acquire health status data of the energy storage device, construct a state feature vector based on the health status data, and use the state feature vector as input data for a deep reinforcement learning method; Based on the state feature vector, the deep reinforcement learning method is used to extract the number of cycles, deep cycle rate, and operating temperature parameters of the energy storage device. Based on the deep reinforcement learning method, a time decay factor is introduced to adaptively weight the operating data at different times in the state feature vector to obtain weighted historical loss data; Analyzing the dynamic change characteristics of the number of cycles, the deep cycle rate, and the operating temperature parameters based on the weighted historical loss data and inputting the dynamic change characteristics into a loss benchmark model to obtain a benchmark loss characteristic of the energy storage device; Based on the benchmark loss characteristics, a multi-dimensional life loss coefficient matrix is ​​constructed using the deep reinforcement learning method, and the multi-dimensional life loss coefficient matrix is ​​optimized by minimizing the deviation between the predicted loss value and the actual loss value to obtain an optimized loss coefficient matrix; The optimized loss coefficient matrix is ​​integrated with time to obtain a cumulative loss state of the energy storage device, and the remaining life of the energy storage device is determined based on the cumulative loss state and the design life of the energy storage device.

5. The method according to claim 4, characterized in that The deep reinforcement learning method is used to construct a multi-dimensional life loss coefficient matrix, and the multi-dimensional life loss coefficient matrix is ​​optimized by minimizing the deviation between the predicted loss value and the actual loss value. The optimized loss coefficient matrix includes: A multi-dimensional life loss coefficient matrix includes a cycle number dimension, a deep cycle rate dimension, and a temperature dimension; minimizing the deviation between the predicted loss value output by the multi-dimensional life loss coefficient matrix and the actual loss value; The multi-dimensional life loss coefficient matrix is ​​optimized based on the result of the minimization process to obtain an optimized loss coefficient matrix.

6. The method according to claim 1, wherein Based on the deep reinforcement learning model, multiple groups of candidate charge and discharge scheduling strategies are generated. The predicted revenue and life loss are used as dual constraints of the reward function. Through alternating training of the policy network and the value network, the generation of charge and discharge strategies that maximize revenue while considering the life of the equipment is achieved. The generation includes: A strategy generation network is constructed based on a deep reinforcement learning model. The strategy generation network includes a parameter generation module and a value assessment module. The network uses the grid load status, electricity price information, device charge status, and ambient temperature of the energy storage device as the system state to generate corresponding charge and discharge action strategy parameters. The system state and the charge and discharge action strategy parameters are combined to form a state-action pair. constructing a dual-constraint reward function based on the state-action pair, inputting the system state into the parameter generation module to obtain a mean parameter and a standard deviation parameter of the action distribution, and inputting the system state and the dual-constraint reward function into the value evaluation module to obtain a state value; An advantage value is calculated based on the state value and the dual-constraint reward function, the advantage value is used to guide the parameter update of the policy generation network, and the temporal difference error of the state value is used to guide the parameter update of the policy generation network, and an optimized policy generation network is obtained through alternating training; The prediction result of the optimized strategy generation network is substituted into the dual-constraint reward function to obtain a strategy evaluation index. Based on the strategy evaluation index, the optimal charge and discharge strategy is selected under the condition of satisfying the life loss constraint to obtain the optimal charge and discharge control strategy of the energy storage device.

7. The method according to claim 6, characterized in that Substituting the prediction result of the optimized strategy generation network into the dual-constraint reward function to obtain a strategy evaluation index, and selecting the optimal charge and discharge strategy based on the strategy evaluation index while satisfying the life loss constraint includes: Substituting the prediction results of the optimized strategy generation network into a dual-constraint reward function, the dual-constraint reward function including a revenue index and a life loss index of the energy storage device, and evaluating the prediction results based on the dual-constraint reward function to obtain a strategy evaluation index corresponding to the prediction results; Calculating a cumulative life loss value of the energy storage device based on the life loss index in the prediction result, and comparing the cumulative life loss value with a preset loss threshold to obtain a life loss constraint condition of the prediction result; The prediction results that meet the life loss constraint condition are sorted in descending order according to the size of the strategy evaluation index, and the results of the descending sorting are used as candidate charging and discharging strategies for the energy storage device; and the charging and discharging strategy with the largest strategy evaluation index among the candidate charging and discharging strategies is used as the optimal charging and discharging control strategy for the energy storage device.

8. An intelligent optimization system for the revenue model of a commercial energy storage power station, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is configured to obtain historical operating data of the commercial energy storage power station, wherein the historical operating data includes charge and discharge power data, battery state of charge data, real-time grid electricity price data, and energy storage device health status data; A second unit is configured to construct a deep reinforcement learning model for a commercial energy storage power station based on the historical operating data, wherein the deep reinforcement learning model adopts a deep neural network structure and nonlinearly maps the historical operating data through the deep neural network structure to generate a predicted profit value; A third unit is configured to analyze the health status data of the energy storage device using the deep reinforcement learning model, establish a multi-dimensional life loss coefficient matrix based on the number of cycles, deep cycle rate, and operating temperature, and implement an assessment of the life loss of the device; A fourth unit is configured to generate multiple sets of candidate charge and discharge scheduling strategies based on the deep reinforcement learning model, using predicted revenue and life loss as dual constraints of the reward function, and to generate charge and discharge strategies that maximize revenue while taking into account equipment life through alternating training of the policy network and the value network; The fifth unit is used to generate an operation control instruction according to the optimized charge and discharge scheduling strategy, and send the operation control instruction to the energy management system of the commercial energy storage power station to control the energy storage power station to operate according to the optimized charge and discharge scheduling strategy.

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

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

Images