Energy storage installed capacity assessment method

Abstract

This invention relates to the field of energy storage technology, specifically to a method for evaluating the installed capacity of energy storage. The method steps of this invention are as follows: S1, collect historical data and perform preprocessing operations on the data; S2, build a dynamic collaborative evaluation model, including: a gap prediction module, a bidirectional attenuation correction module, a multi-energy flow coupling module, and a multi-objective optimization module; S3, the dynamic collaborative evaluation model obtains a feasible installed capacity range through data transfer and reasoning between modules; S4, determine the final scheme according to the required scenario. The beneficial effects of this invention are: (1) the power gap prediction model based on LSTM neural network can achieve accurate prediction of power gap; (2) the introduction of PSO algorithm to adjust the LSTM layer parameters improves the long-term accuracy of the prediction results; (3) in the evaluation, the introduction of device degradation and repair factors can improve the comprehensiveness of the evaluation.

Description

Technical Field

[0001] This invention relates to the field of energy storage technology, and more specifically to a method for evaluating the installed capacity of energy storage. Background Technology

[0002] As a guarantee for balancing energy supply and demand, energy storage systems are becoming increasingly important. The assessment of energy storage installed capacity is not only a prerequisite for the planning and construction of energy storage projects, but also a key link in achieving efficient, stable and economical operation of the energy system.

[0003] Currently, most common assessment methods focus on a single scenario and primarily assess current capacity, without considering long-term capacity requirements. Summary of the Invention

[0004] The main objective of this invention is to provide a method for evaluating the installed capacity of energy storage, which introduces attenuation coefficient and recovery coefficient, takes into account the aging and repair factors of the device, and solves the problems in the background art mentioned above.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an energy storage installed capacity assessment method, comprising the following steps: S1. Collect historical data and perform preprocessing operations on the data; S2. Build a dynamic collaborative evaluation model, including: a gap prediction module, a bidirectional attenuation correction module, a multi-energy flow coupling module, and a multi-objective optimization module; S3, the dynamic collaborative evaluation model obtains feasible installed capacity ranges through data transfer and reasoning between modules; S4. Determine the final solution based on the required scenario.

[0006] Furthermore, the historical data includes: energy parameters, load parameters, energy storage system parameters, operating parameters, policy parameters, grid parameters, and climate parameters.

[0007] Furthermore, in step S3, the transmission and reasoning process is as follows: the gap prediction module obtains the predicted load and output predicted power curves; The multi-energy flow coupling module determines the power gap based on load and output power prediction data; The bidirectional attenuation correction module corrects the available energy capacity of the energy storage system throughout its entire life cycle based on the power gap; The multi-objective optimization module calculates the optimal capacity range, i.e. the feasible installed capacity range, based on the constraints. Furthermore, the gap prediction module includes a power gap prediction model based on an LSTM neural network; The model includes: an input layer, an LSTM layer, an attention layer, and an output layer; The input layer takes multi-dimensional feature parameters as model input, and the output layer outputs the predicted power output and load power of each type of energy. The LSTM layer has a two-layer structure with the same structure, consisting of LSTM1 and LSTM2 layers, including forget gate, input gate, and output gate; The forgetting gate uses the sigmoid function to determine the degree of forgetting of cell state information from the previous time step. The process is as follows: ; in, f ( t ) is the output vector of the forget gate. W f , b f These are the weights and biases of the forget gate, respectively. h t-1 The hidden state retained from the previous moment. x t This is the input for the current moment; The input gate uses the tanh function to determine what information needs to be added to the cell state at the current time step. The process is as follows: ; in, i ( t ) represents the output vector of the input gate. W i , b i These are the weights and biases of the input gate, respectively; And generate candidate memory vectors: ; in, Candidate memory vectors, W C , b C These are the weights and biases of the candidate memories, respectively. Update the cell state by combining the outputs of the forget gate and the input gate: ; in, C t In cellular state, The symbol for element-wise multiplication; The output gate determines which information from the current memory cell state needs to be output to the hidden state: ; ; in, h tFor the new hidden state, O ( t ) represents the output vector of the output gate. W O , b O These are the weights and biases of the forget gate, respectively; The attention layer calculates the importance score of the hidden state at each time step. e t : ; in, u For learnable parameter vectors, W h , b h These are the weights and biases of the attention layer, respectively. Based on the importance score, determine the attention weight for each time step. : ; in, T 0 represents the total number of steps; Then the context vector D This can be represented using attention weights: ; The output layer outputs the future first t Predicted value of step : ; in, W y , b y These are the weight matrix and bias vector of the output layer, respectively; It is composed of the context vector and the hidden state of the last time step.

[0008] Furthermore, the training process of the prediction model is as follows: S501. Set the initial values ​​of the hyperparameters, which include: number of samples per batch, maximum number of training rounds, and initial learning rate; S502. Divide the 3D data into two sets according to the proportions: training set and validation set, and use the training set as the input to the model. S503. Determine the initial values ​​of each weight and bias in the LSTM neural network; S504. Train the training set samples in batches according to the set number of samples in each batch to obtain the output value; The output values ​​include: the predicted total load and the predicted total output. S505. After each training batch, calculate the output loss value based on the determined loss function: ; in, L It is the loss value. For the first t The actual value of the step; S506. After a batch of data has been trained, the weights and biases of each gate are updated using gradient descent. S507. Using the validation set as input, repeat steps S504-S505 to obtain the loss value and verify the accuracy of the model. S508. Repeat steps S504 to S507 until the number of training iterations is greater than the minimum number of training iterations. At this point, judge the loss value obtained in S507. When the loss value is less than or equal to the cutoff threshold, the training ends early. Training ends when the maximum number of repetitions is reached.

[0009] Furthermore, the prediction model will provide feedback to the model based on the prediction results, and the process is as follows: First, calculate the deviation coefficients corresponding to the load power and output power. Both satisfy the deviation coefficient formula: ; in, Indicates the deviation coefficient; When either the load power deviation coefficient or the output power deviation coefficient exceeds a threshold, the PSO algorithm is used to optimize the LSTM layer weights and bias parameters in the LSTM neural network. The optimization process is as follows: S601. Randomly generate several particles and randomly initialize the particles within their limited range; The current position of each particle Determined as the individual's optimal position Pb i ,Will Pb i The position with the highest fitness is defined as the global optimum. Gb ; S602. Determine the fitness function based on the deviation coefficient: ; Where F is the fitness value, , These are the load power deviation coefficient and the output power deviation coefficient, respectively. S603, Update the particle velocity and position: ; in, , Let be the velocity and position of the i-th particle in the k-th iteration, and w be the inertial weight. , These are learning factors, , These are uniformly random numbers; When the particle velocity and position exceed the boundary, update the particle position on the boundary and update the particle velocity in the opposite direction: ; Where n is the adjustment coefficient, which is always positive; S604. Calculate the fitness of the new position. If the fitness of the new position is greater than the fitness of the current position, then update the new position. Pb i ; Will Pb i The position with the highest fitness is defined as the global optimum. Gb ; S605. Repeat steps S603 to S604. When the number of iterations is reached or the early stopping condition is met, stop training and update the parameters in the global optimal position to the parameters of the LSTM layer.

[0010] Furthermore, the multi-energy flow coupling module establishes a multi-energy flow cooperative equation based on the predicted output power and predicted load power output from the prediction model: ; in, For the total power shortfall, For power shortage, , These represent the power deficit of other forms of energy and their corresponding energy conversion coefficients; the power deficit is the difference between load power and output power. Use vehicle-to-everything (V2X) interaction correction factor Adjustments to energy storage demand: ; in, This is the corrected power gap.

[0011] Furthermore, the bidirectional attenuation correction module, after considering the device's losses, adjusts its actual usable capacity. Make corrections: ; in, For the initial capacity, , These are the cyclic decay coefficient and the time decay coefficient, respectively. , These are the cumulative number of iterations and the rated number of iterations, respectively. , These are cumulative runtime and designed operating years, respectively. It is the environmental correction factor. It is the first Capacity restoration coefficient for each maintenance. It is the repair aging factor; the initial capacity can be obtained by integrating the corrected power gap; The module also measures charge and discharge efficiency. Corrections have been made: ; in, This represents the initial value for charge / discharge efficiency. It is the efficiency attenuation coefficient. It is the first Efficiency restoration coefficient of the next maintenance.

[0012] Furthermore, the multi-objective optimization module constructs a three-dimensional objective function, including: economic objective and stability objective; The economic objective function is: ; in, For total cost, , These are the construction costs per unit of energy and per unit of power, respectively. The discount rate is... , These are the maintenance costs per unit of energy and per unit of power, respectively. , These are the unit prices for the deficit penalty per unit of energy and per unit of power, respectively. , These are the total annual energy and power deficits, respectively. It is a small value to prevent the denominator from being 0; The stability objective function is: ; in, For stability, Rated power, The difference between the actual power and the rated power. For the evaluation period; The economy and stability of energy use are subject to the following constraints: From an economic perspective, constraints are imposed based on the principle that marginal construction and maintenance costs equal marginal penalty costs. The upper and lower limits of these economic constraints are determined as follows: ; in, , These are the lower and upper limits of energy economic constraints, respectively. From a stability perspective, constraints are applied: the upper and lower limits of stability constraints are determined as follows: ; in, , These are the lower and upper limits of the energy stability constraint, respectively. , These are the minimum and maximum safe energy values ​​for energy storage. , These are the lower and upper limits allowed for the SOC value, respectively. Energy capacity The range of values ​​for is: ; The economy and stability of power output are subject to the following constraints: From an economic perspective, the upper and lower limits of the economic constraints are determined as follows: ; in, , These are the lower and upper limits of the power economy constraint, respectively; From a stability perspective, constraints are applied: the upper and lower limits of stability constraints are determined as follows: ; in, , These are the lower and upper limits of the power stability constraint, respectively. , These are the output power fluctuation rate and the load power fluctuation rate, respectively. , , These are the inverter's rated power, grid connection limit, and maximum battery discharge power, respectively. Power capacity The range of values ​​for is: .

[0013] Furthermore, in step S4, the final solution is determined according to different scenarios: when economy is the priority, the energy capacity and power capacity corresponding to the lowest total cost are taken as the final solution. When stability is the priority, the energy capacity and power capacity corresponding to the highest stability are taken as the final solution.

[0014] Beneficial effects: (1) The power gap prediction model based on LSTM neural network can achieve accurate prediction of power gap; (2) Introduce the PSO algorithm to adjust the parameters of the LSTM layer and improve the long-term accuracy of the prediction results; (3) Introducing the factors of device degradation and repair during the assessment can improve the comprehensiveness of the assessment. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0016] Example 1 like Figure 1 The method for evaluating energy storage installed capacity, as shown in the figure, comprises the following steps: S1. Collect historical data and perform preprocessing operations on the data; The data includes: energy parameters, load parameters, energy storage system parameters, operating parameters, policy parameters, grid parameters, and climate parameters; This includes: photovoltaic / wind power output forecast curves, thermal power peak-shaving rate, hydropower curtailment rate, and hydrogen energy storage conversion efficiency; Load parameters include: real-time load data and electric vehicle charging load curves; The parameters of the energy storage system itself include: initial charge and discharge efficiency, rated cycle life, unit cost, maximum charge and discharge power, aging degradation coefficient matrix, design operating years, and maintenance and repair coefficient. Policy parameters include: per-kilowatt-hour subsidies, carbon prices, new energy consumption assessment indicators, and carbon tax policies; Power grid parameters include: transmission line capacity, node voltage constraints, short-circuit capacity, and network loss rate; Climate parameters, including: light intensity, wind speed probability distribution, and extreme weather warning data; Operating parameters include: cumulative number of cycles, cumulative runtime, ambient temperature, and real-time maintenance records; The aging degradation coefficient matrix is ​​the ratio of capacity to efficiency degradation rate.

[0017] S2. Build a dynamic collaborative evaluation model, including: a gap prediction module, a bidirectional attenuation correction module, a multi-energy flow coupling module, and a multi-objective optimization module; The gap prediction module includes a power gap prediction model based on an LSTM neural network; The model includes: an input layer, an LSTM layer, an attention layer, and an output layer; The input layer uses the feature factors extracted from the data in step S1 as the model input, including: historical photovoltaic power generation data, historical wind power generation data, historical industrial load data, historical residential load data, historical regional solar irradiance data, historical regional wind speed data, historical ambient temperature data, and time feature data. The time feature data will be marked with codes such as weekends, weekdays, and holidays. The LSTM layer has a two-layer structure with the same structure, only the number of neurons is different. It is divided into LSTM1 layer and LSTM2 layer. LSTM1 layer contains 128 neurons and LSTM2 layer contains 64 neurons, including forget gate, input gate and output gate. The forgetting gate uses the sigmoid function to determine the degree of forgetting of cell state information from the previous time step. The process is as follows: ; in, f ( t ) is the output vector of the forget gate. W f , b f These are the weights and biases of the forget gate, respectively. h t-1 The hidden state retained from the previous moment. x t This is the input for the current moment; The input gate uses the tanh function to determine what information needs to be added to the cell state at the current time step. The process is as follows: ; in, i ( t ) represents the output vector of the input gate. W i , b i These are the weights and biases of the input gate, respectively; And generate candidate memory vectors: ; in, Candidate memory vectors, W C , b C These are the weights and biases of the candidate memories, respectively. Update the cell state by combining the outputs of the forget gate and the input gate: ; in, C t In cellular state, The symbol for element-wise multiplication; The output gate determines which information from the current memory cell state needs to be output to the hidden state: ; ; in, h t For the new hidden state, O ( t ) represents the output vector of the output gate. W O , b O These are the weights and biases of the forget gate, respectively; The attention layer calculates the importance score of the hidden state at each time step. e t : ; in, u For learnable parameter vectors, W h , b h These are the weights and biases of the attention layer, respectively. Based on the importance score, determine the attention weight for each time step. : ; in, T 0 represents the total number of steps; Then the context vector D This can be represented using attention weights: ; The output layer outputs predicted power output and load power for each type of energy source, showing the future output of the [missing information]. t Predicted value of step : ; in, W y , b y These are the weight matrix and bias vector of the output layer, respectively; It is composed of the context vector and the hidden state of the last time step.

[0018] The training process of the prediction model is as follows: A1. Set the initial values ​​of the hyperparameters, which include: 64 samples per batch, 200 maximum training epochs, and an initial learning rate of 0.005. A2. Divide the 3D data into a training set and a validation set in an 8:2 ratio, and use the training set as the input to the model; A3. Determine the initial values ​​of each weight and bias in the LSTM neural network; the weights and biases are based on... A4. Train the training set samples in batches according to the set number of samples in each batch to obtain the output value; The output values ​​include: the predicted total load and the predicted total output. A5. After each training batch, calculate the output loss value based on the determined loss function: ; in, L It is the loss value. For the first t The actual value of the step; A6. After training on a batch of data, use gradient descent to update the weights and biases of each gate: (25); in, These are the updated values ​​for the weights and biases. For the weights and biases of the hidden or output layer, The learning rate; A7. Using the validation set as input, repeat steps A4 to A5 to obtain the loss value and verify the accuracy of the model. A8. Repeat steps A4 to A7 until the number of training iterations is greater than the minimum number of training iterations. At this point, judge the loss value obtained in A7. When the loss value is less than or equal to the cutoff threshold, the training ends early. Training ends when the maximum number of repetitions is reached.

[0019] The predictive model also provides feedback based on the prediction results; the process is as follows: First, calculate the deviation coefficients corresponding to the load power and output power. Both satisfy the deviation coefficient formula: ; in, Indicates the deviation coefficient; When either the load power deviation coefficient or the output power deviation coefficient exceeds a threshold, the PSO algorithm is used to optimize the LSTM layer weights and bias parameters in the LSTM neural network. The optimization process is as follows: B1. Randomly generate 50 particles, including weight parameters, bias parameters, and velocities, and randomly initialize the particles within a defined range; the weight parameter range is... The bias parameter range is The speed range is ; The current position of each particle Determined as the individual's optimal position Pb i ,Will Pb i The position with the highest fitness is defined as the global optimum. Gb ; B2. Determine the fitness function based on the deviation coefficient: ; Where F is the fitness value, , These are the load power deviation coefficient and the output power deviation coefficient, respectively. B3. Update the particle velocity and position: ; in, , Let be the velocity and position of the i-th particle in the k-th iteration, and w be the inertial weight. , These are the learning factors; a value of 2 yields better results. , These are uniformly random numbers, with a value range of [value range missing]. ; When the particle velocity and position exceed the boundary, update the particle position on the boundary and update the particle velocity in the opposite direction: ; Where n is the adjustment coefficient, which is always positive; B4. Calculate the fitness of the new position. If the fitness of the new position is greater than the fitness of the current position, then update the new position. Pb i ; Will Pb i The position with the highest fitness is defined as the global optimum. Gb ; B5. Repeat steps B3 to B4. When the number of iterations is reached or the early stopping condition is met, stop training and update the parameters in the global optimal position to the parameters of the LSTM layer.

[0020] The multi-energy flow coupling module establishes multi-energy flow cooperative equations based on the predicted power output values ​​of various energy types and the predicted load power output values ​​output by the gap prediction module: ; in, For the total power shortfall, For power shortage, , These represent the power gaps of other forms of energy and their corresponding energy conversion coefficients; where each power gap value is equal to the difference between the predicted load power and the predicted output power. Use vehicle-to-everything (V2X) interaction correction factor Adjustments to energy storage demand: ; in, This is the corrected power gap.

[0021] The bidirectional attenuation correction module, after considering the device's losses, adjusts its actual usable capacity. Make corrections: ; in, For the initial capacity, , These are the cyclic decay coefficient and the time decay coefficient, respectively. , These are the cumulative number of iterations and the rated number of iterations, respectively. , These are cumulative runtime and designed operating years, respectively. It is the environmental correction factor. It is the first Capacity restoration coefficient for each maintenance. It is the repair time factor; the initial capacity can be obtained by integrating the corrected power gap; The module also measures charge and discharge efficiency. Corrections have been made: ; in, This represents the initial value for charge / discharge efficiency. It is the efficiency attenuation coefficient. It is the first The efficiency recovery coefficient of the next maintenance; The required energy capacity has now been determined. The reference value is the actual usage capacity. Power capacity The reference value is the corrected power gap. .

[0022] The multi-objective optimization module constructs a three-dimensional objective function, including: economic objective and stability objective; The economic objective function is: ; in, For total cost, , These are the construction costs per unit of energy and per unit of power, respectively. The discount rate is... , These are the maintenance costs per unit of energy and per unit of power, respectively. , These are the unit prices for the deficit penalty per unit of energy and per unit of power, respectively. , These are the total annual energy and power deficits, respectively. It is a small value to prevent the denominator from being 0; The stability objective function is: ; in, For stability, Rated power, The difference between the actual power and the rated power. For the evaluation period; The economy and stability of energy use are subject to the following constraints: From an economic perspective, constraints are imposed based on the principle that marginal construction and maintenance costs equal marginal penalty costs. The upper and lower limits of these economic constraints are determined as follows: ; in, , These are the lower and upper limits of energy economic constraints, respectively. From a stability perspective, constraints are applied: the upper and lower limits of stability constraints are determined as follows: ; in, , These are the lower and upper limits of the energy stability constraint, respectively. , These are the minimum and maximum safe energy values ​​for energy storage. , These are the lower and upper limits allowed for the SOC value, respectively. Energy capacity The range of values ​​for is: ; The economy and stability of power output are subject to the following constraints: From an economic perspective, the upper and lower limits of the economic constraints are determined as follows: ; in, , These are the lower and upper limits of the power economy constraint, respectively; From a stability perspective, constraints are applied: the upper and lower limits of stability constraints are determined as follows: ; in, , These are the lower and upper limits of the power stability constraint, respectively. , These are the output power fluctuation rate and the load power fluctuation rate, respectively. , , These are the inverter's rated power, grid connection limit, and maximum battery discharge power, respectively. Power capacity The range of values ​​for is: .

[0023] S3, the dynamic collaborative evaluation model obtains solutions applicable to different scenarios through data transfer and reasoning between modules; Its transmission and reasoning process is as follows: the gap prediction module obtains the predicted load and output power curves; The multi-energy flow coupling module determines the power gap based on load and output power prediction data; The bidirectional attenuation correction module corrects the available energy capacity of the energy storage system throughout its entire life cycle based on the power gap; The multi-objective optimization module calculates the optimal capacity range, i.e. the feasible installed capacity range, based on the constraints.

[0024] S4. Determine the final solution based on different scenarios: When economy is the priority, the energy capacity and power capacity corresponding to the lowest total cost are taken as the final solution; when stability is the priority, the energy capacity and power capacity corresponding to the highest stability are taken as the final solution.

[0025] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for evaluating the installed capacity of energy storage, characterized in that: Includes the following steps: S1. Collect historical data and perform preprocessing operations on the data; S2. Build a dynamic collaborative evaluation model, including: a gap prediction module, a bidirectional attenuation correction module, a multi-energy flow coupling module, and a multi-objective optimization module; S3, the dynamic collaborative evaluation model obtains feasible installed capacity ranges through data transfer and reasoning between modules; The transmission and reasoning process is as follows: the gap prediction module obtains the predicted load and output power curves; The multi-energy flow coupling module determines the power gap based on load and output power prediction data; The bidirectional attenuation correction module corrects the available energy capacity of the energy storage system throughout its entire life cycle based on the power gap; The multi-objective optimization module calculates the optimal capacity range, i.e. the feasible installed capacity range, based on the constraints. The bidirectional attenuation correction module, after considering the device's losses, adjusts its actual usable capacity. Make corrections: (1); in, For the initial capacity, , These are the cyclic decay coefficient and the time decay coefficient, respectively. , These are the cumulative number of iterations and the rated number of iterations, respectively. , These are cumulative runtime and designed operating years, respectively. It is the environmental correction factor. It is the first Capacity restoration coefficient for each maintenance. It is a repair time factor. The total number of maintenance cycles; the initial capacity can be obtained by integrating the corrected power gap. The module also measures charge and discharge efficiency. Corrections have been made: （2）； in, This represents the initial value for charge / discharge efficiency. It is the efficiency attenuation coefficient. It is the first The efficiency recovery coefficient of the next maintenance; S4. Determine the final solution according to the required scenario. Specifically, when economy is the priority, the energy capacity and power capacity corresponding to the lowest total cost shall be the final solution; when stability is the priority, the energy capacity and power capacity corresponding to the highest stability shall be the final solution.

2. The evaluation method according to claim 1, characterized in that: The historical data includes: energy parameters, load parameters, energy storage system parameters, operating parameters, policy parameters, grid parameters, and climate parameters.

3. The evaluation method according to claim 1, characterized in that: The gap prediction module includes a power gap prediction model based on an LSTM neural network; The model includes: an input layer, an LSTM layer, an attention layer, and an output layer; The input layer takes multi-dimensional feature parameters as model input, and the output layer outputs the predicted power output and load power of each type of energy. The LSTM layer has a two-layer structure with the same structure, consisting of LSTM1 and LSTM2 layers, including forget gate, input gate, and output gate; The forgetting gate uses the sigmoid function to determine the degree of forgetting of cell state information from the previous time step. The process is as follows: （3）； in, This is the output vector of the forget gate. , These are the weights and biases of the forget gate, respectively. The hidden state retained from the previous moment. This is the input for the current moment; The input gate uses the tanh function to determine what information needs to be added to the cell state at the current time step. The process is as follows: （4）； in, The output vector of the input gate. , These are the weights and biases of the input gate, respectively; And generate candidate memory vectors: （5）； in, Candidate memory vectors, , These are the weights and biases of the candidate memories, respectively. Update the cell state by combining the outputs of the forget gate and the input gate: （6）； in, In cellular state, The symbol for element-wise multiplication; The output gate determines which information from the current memory cell state needs to be output to the hidden state: （7）； （8）； in, For the new hidden state, The output vector of the output gate. , These are the weights and biases of the output gate, respectively. The attention layer calculates the importance score of the hidden state at each time step. : （9）； in, For learnable parameter vectors, , These are the weights and biases of the attention layer, respectively. Based on the importance score, determine the attention weight for each time step. : （10）； in, T 0 represents the total number of steps; Then the context vector D This can be represented using attention weights: （11）； The output layer outputs the future first t Predicted value of step : （12）； in, , These are the weight matrix and bias vector of the output layer, respectively; It is composed of the context vector and the hidden state of the last time step.

4. The evaluation method according to claim 3, characterized in that: The training process of the prediction model is as follows: S501. Set the initial values ​​of the hyperparameters, which include: number of samples per batch, maximum number of training rounds, and initial learning rate; S502. Divide the 3D data into two sets according to the proportions: training set and validation set, and use the training set as the input to the model. S503. Determine the initial values ​​of each weight and bias in the LSTM neural network; S504. Train the training set samples in batches according to the set number of samples in each batch to obtain the output value; The output values ​​include: the predicted total load and the predicted total output. S505. After each training batch, calculate the output loss value based on the determined loss function: （13）； in, L It is the loss value. The total number of samples in each training batch. For the first The actual value of the step; S506. After a batch of data has been trained, the weights and biases of each gate are updated using gradient descent. S507. Using the validation set as input, repeat steps S504-S505 to obtain the loss value and verify the accuracy of the model. S508. Repeat steps S504 to S507 until the number of training iterations is greater than the minimum number of training iterations. At this point, judge the loss value obtained in S507. When the loss value is less than or equal to the cutoff threshold, the training ends early. Training ends when the maximum number of repetitions is reached.

5. The evaluation method according to claim 4, characterized in that: The prediction model will provide feedback based on the prediction results, and the process is as follows: First, calculate the deviation coefficients corresponding to the load power and output power. Both satisfy the deviation coefficient formula: （14）； in, Indicates the deviation coefficient; When either the load power deviation coefficient or the output power deviation coefficient exceeds a threshold, the PSO algorithm is used to optimize the LSTM layer weights and bias parameters in the LSTM neural network. The optimization process is as follows: S601. Randomly generate several particles and randomly initialize the particles within their limited range; The current position of each particle Determined as the individual's optimal position ,Will The position with the highest fitness is defined as the global optimum. ; S602. Determine the fitness function based on the deviation coefficient: （15）； in, For fitness value, , These are the load power deviation coefficient and the output power deviation coefficient, respectively. S603, Update the particle velocity and position: （16）； in, , For the first The particle in the first Velocity and position at the next iteration For inertial weights, , These are learning factors, , These are uniformly random numbers; When the particle velocity and position exceed the boundary, update the particle position on the boundary and update the particle velocity in the opposite direction: （17）； in, The adjustment coefficient is always positive. S604. Calculate the fitness of the new position. If the fitness of the new position is greater than the fitness of the current position, then update the new position. ; Will The position with the highest fitness is defined as the global optimum. ; S605. Repeat steps S603 to S604. When the number of iterations is reached or the early stopping condition is met, stop training and update the parameters in the global optimal position to the parameters of the LSTM layer.

6. The evaluation method according to claim 4, characterized in that: The multi-energy flow coupling module establishes a multi-energy flow cooperative equation based on the predicted output power and predicted load power output from the prediction model: （18）； in, For the total power shortfall, For power shortage, , These represent the power deficit of other forms of energy and their corresponding energy conversion coefficients, respectively. This represents the total number of other forms of energy; the power gap is the difference between load power and output power. Use vehicle-to-everything (V2X) interaction correction factor Adjustments to energy storage demand: （19）； in, This is the corrected power gap.

7. The evaluation method according to claim 1, characterized in that: The multi-objective optimization module constructs a three-dimensional objective function, including: an economic objective and a stability objective; The economic objective function is: （20）； in, For total cost, For energy capacity, For power capacity, , These are the construction costs per unit of energy and per unit of power, respectively. The discount rate is... , These are the maintenance costs per unit of energy and per unit of power, respectively. , These are the unit prices for the deficit penalty per unit of energy and per unit of power, respectively. , These are the total annual energy and power deficits, respectively. It is a small value to prevent the denominator from being 0; The stability objective function is: （21）； in, For stability, Rated power, The difference between the actual power and the rated power. For the evaluation period; The economy and stability of energy use are subject to the following constraints: From an economic perspective, constraints are imposed based on the principle that marginal construction and maintenance costs equal marginal penalty costs. The upper and lower limits of these economic constraints are determined as follows: （22）； in, , These are the lower and upper limits of energy economic constraints, respectively. From a stability perspective, constraints are applied: the upper and lower limits of stability constraints are determined as follows: （23）； in, , These are the lower and upper limits of the energy stability constraint, respectively. , These are the minimum and maximum safe energy values ​​for energy storage. , These are the lower and upper limits allowed for the SOC value, respectively. Energy capacity The range of values ​​for is: （24）； The economy and stability of power output are subject to the following constraints: From an economic perspective, the upper and lower limits of the economic constraints are determined as follows: （25）； in, , These are the lower and upper limits of the power economy constraint, respectively; From a stability perspective, constraints are applied: the upper and lower limits of stability constraints are determined as follows: （26）； in, , These are the lower and upper limits of the power stability constraint, respectively. , These are output power and load power, respectively. , , These are the inverter's rated power, grid connection limit, and maximum battery discharge power, respectively. Power capacity The range of values ​​for is: （27）。

Images