Economical operation control method and system for distributed energy storage and new energy consumption

Abstract

The invention discloses an economic operation control method and system for distributed energy storage and new energy consumption, relates to the technical field of energy storage, and aims to acquire meteorological prediction and historical data of wind power and photoelectric locations and establish a meteorological factor time sequence through data collection and preprocessing. And then, a hybrid prediction model is constructed based on a physical principle model in combination with a data-driven machine learning model to predict the generating capacity. And then determining charging decision limitation of the energy storage system according to a power generation amount prediction value and future power consumption prediction, further determining a charging and discharging decision function, and adjusting charging and discharging power by considering a market electricity price factor. And a mathematical model is constructed, and a gradient descent optimization algorithm is utilized to optimize charging parameters considering the service life of the battery. Energy storage charging and discharging operation can be better planned, economic benefits are improved, the service life of the battery is effectively prolonged, and the system can better adapt to different conditions and keep good performance.

Description

Technical Field

[0001] The present invention relates to the field of energy storage technology, and in particular to an economic operation control method and system for absorbing new energy through distributed energy storage. Background Art

[0002] In the era of energy transformation, new energy power generation is booming, and wind power and photovoltaic power generation have attracted much attention. However, their intermittent and volatile characteristics have brought severe challenges to the power consumption of the power system. Wind power generation is affected by wind speed, and the instability of wind speed causes large fluctuations in power generation; photovoltaic power generation is closely related to the intensity of solar radiation, and weather changes can easily change its power generation; as a result, it is difficult to accurately match new energy power with traditional power load demand, and excess or insufficient power often occurs, seriously interfering with the supply and demand balance and power quality of the power system. The emergence of energy storage technology has brought hope for solving this problem. Distributed energy storage systems can store electricity when there is excess new energy power generation and release electricity when there is insufficient power, playing the role of "peak shaving and valley filling" and enhancing the stability and reliability of the power system.

[0003] However, there are still many shortcomings in the current relevant technologies. In terms of power generation forecasting, although the traditional model based on physical principles can calculate the theoretical power generation according to relevant theories and parameters and provide a basic framework for forecasting, it is difficult to accurately capture the nonlinear changes in power generation caused by complex and changeable meteorological conditions and equipment performance changes in actual operation. Although the data-driven machine learning model can mine the nonlinear relationships and complex patterns in the data to improve the accuracy of prediction, it is highly dependent on the quality and quantity of data, and the model has weak interpretability. In the control of energy storage systems, the problem is also prominent. When determining the charging decision limit, the detailed historical data and the accurate prediction of future electricity consumption have not been fully combined in the past, resulting in the energy storage system being unable to charge at the best time and unable to effectively store or release electricity. When determining the charging and discharging power, most of them only consider the power generation, electricity consumption and the power status of the energy storage system, ignoring the market electricity price factor. Under the dynamic changes in the electricity market, this makes it difficult for the energy storage system to maximize its economic benefits and miss the profit opportunities brought by electricity price fluctuations. In addition, battery life has not been given sufficient attention in existing control strategies, and the charging process does not fully consider the impact of charging current, charging depth, and charging temperature on battery life, resulting in premature aging of the battery, performance degradation, increased operation and maintenance costs, and reduced overall economic benefits.

[0004] In view of the above limitations, in order to more efficiently realize the consumption of new energy by distributed energy storage and improve the economic operation efficiency of the system, an economic operation control method and system for distributed energy storage to consume new energy is proposed. A hybrid model is constructed by integrating multiple prediction algorithms to improve the accuracy of power generation prediction, and reasonable charging decision limits are determined accordingly. At the same time, the charging and discharging power is adjusted in combination with the market electricity price. The gradient descent optimization algorithm is used to optimize the charging parameters related to battery life to extend the battery life, thereby promoting the economic, efficient and stable operation of the distributed energy storage system when absorbing new energy. Summary of the invention

[0005] The purpose of the present invention is to provide an economic operation control method and system for distributed energy storage to absorb new energy. By collecting future forecast data from local meteorological departments and comparing historical data, a forecasting mathematical algorithm is established. A hybrid model is constructed by combining a variety of different types of forecasting algorithms to predict the power generation of energy generation equipment. The predicted power generation is used to plan the charging and discharging operations of energy storage in advance to better absorb new energy.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method and system for controlling the economic operation of distributed energy storage and absorbing new energy, characterized in that it includes the following steps:

[0008] S1: Data collection and preprocessing: Obtain the meteorological forecast data of wind power and photovoltaic power stations from the local meteorological departments, and collect the corresponding historical meteorological data of the past few years to establish the data time series of meteorological factors;

[0009] Obtain the weather forecast data of wind power and photovoltaic power stations from the local meteorological department, collect the corresponding historical weather data of the past few years, and establish the time series data of various meteorological factors; specifically, wind speed v next 、Solar radiation intensity I next , Temperature T next 、Weather conditions next , and collect historical meteorological data from the past few years, which are recorded as wind speed series {v i}、Solar irradiance intensity sequence {I i}、Temperature series {T i}、Weather condition sequence {W n}, where i represents the time index.

[0010] The weather conditions W next It is determined by assigning values ​​to the specific weather conditions, specifically, the clear value is 1, cloudy is 2, overcast is 3, light rain is 4, moderate rain and above is 5, snow is 6, fog is 7, strong wind is 8, and thunderstorm is 9.

[0011] S2: Prediction of wind turbine power generation: Based on the physical principle model, the theoretical power output is calculated according to the aerodynamic principle combined with air density, blade swept area, and wind energy utilization coefficient. The machine learning model based on the long short-term memory network (LSTM) data is used to drive the feature vector containing wind speed, temperature, weather conditions and their changes. The wind power generation is predicted through LSTM unit related operations and fully connected layer mapping steps, and the weight matrix and bias vector are optimized through training. A hybrid prediction model is constructed, and the prediction results of the physical principle model and the LSTM model are integrated by weighted average to obtain the final wind turbine hourly power generation prediction value.

[0012] Step S2 predicts the power generation of the wind turbine generator. Specifically, according to the aerodynamic principle of wind power generation, the theoretical power output P of the wind turbine generator is wtheory (v) is calculated by the following formula:

[0013]

[0014] Where ρ is the air density, A is the swept area of ​​the wind turbine blade, C p is the wind energy utilization coefficient, v is the wind speed;

[0015] The LSTM data-driven machine learning model selects feature vectors from historical data and selects W i The closest historical time point i with the same weather and the same value. For each time point i, the eigenvector X i Including wind speed v i , Temperature T i 、Weather conditions n and the changes in wind speed, temperature, and weather conditions in the previous h hours; the corresponding target value is the actual wind turbine power generation Y at that time point i ;

[0016] Assume that the input dimension of the LSTM network is d, that is, the feature vector X i The dimension of the hidden layer is m, the output dimension is 1, and the LSTM unit formula is as follows:

[0017] Forget Gate:

[0018] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0019] Among them, f t is the output of the forget gate; σ is the sigmoid function, which maps the output value to between 0 and 1, W fis the weight matrix of the forget gate, with dimension (m+d)×m, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous moment, with dimension m, X t is the input feature vector at the current moment, with dimension d;

[0020] Set up the input gate:

[0021] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0022]

[0023] Among them, i t is the output of the input gate, is the new candidate cell state, W i , W C are the corresponding weight matrices with dimensions of (m+d)×m and (m+d)×m, respectively, b i , b C is the corresponding bias vector;

[0024] Set the cell status update:

[0025]

[0026] Among them, C t is the updated cell state, (f t ·C t-1 ) represents the old information that the forget gate decides to forget and represents the new information that the input gate decides to update, C t-1 is the cell state at the previous moment;

[0027] Set up the output gate:

[0028] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0029] h t =o t tanh(C t )

[0030] Among them, t is the output of the output gate, h t is the hidden state at the current moment, W ois the weight matrix of the output gate, with dimension (m+d)×m, b o is the bias vector of the output gate;

[0031] After setting to N time steps, N corresponds to the length of the time series in the historical data, the hidden state h is transformed through a fully connected layer. N Mapped to the predicted power generation value, the weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Y pred for:

[0032] Y pred =W out ·h N +b out

[0033] where Y pred Represents the predicted power generation value, W out is the weight matrix of the fully connected layer, W out This is the weight matrix of the fully connected layer, b out is the bias vector of the fully connected layer;

[0034] Through training data, that is, feature vector X i and target value Y i For all weight matrices (W f , W i , W C , W o , W out ) and the bias vector (b f , b i , b C , b o , b o , b out ) is optimized using the stochastic gradient descent (SGD) optimization algorithm so that the predicted value Y pred As close as possible to the actual power generation value Y i ;

[0035] Combining the above physical principle model with the data-driven machine learning model, a hybrid prediction model is constructed. The final predicted hourly power generation of the wind turbine on the second day is P w (t), the prediction results of the two models are fused by weighted average:

[0036] P w (t) = αP wtheory (v next (t))+(1-α)Y pred (t)

[0037] Among them, t represents the hour index, ranging from 1 to 24, α is the weight coefficient, ranging from 0 to 1, and is set to 0.5 in the initial stage and adjusted according to the prediction effect, v next (t) is the predicted wind speed at hour t of the next day, Y pred (t) is the power generation at the t hour of the next day predicted by the LSTM network.

[0038] S3: Photovoltaic panel power generation prediction: Based on the physical principle model, the theoretical power generation is calculated according to the photovoltaic power generation principle combined with the solar radiation intensity, installation area, conversion efficiency, and system comprehensive efficiency parameters. Through the data-driven machine learning model based on the convolutional neural network (CNN), the feature vectors including solar radiation intensity, temperature, weather conditions and their changes are selected, and the power generation is predicted through the convolution layer, pooling layer processing and fully connected layer mapping operations, and the relevant parameters are trained and optimized; a hybrid prediction photovoltaic panel model is constructed to calculate the final hourly power generation prediction value of the photovoltaic panel by weighted averaging the prediction results of the physical principle model and the CNN model;

[0039] Step S3 predicts the power generation of the photovoltaic panel. Specifically, according to the principle of photovoltaic power generation, the power generation E of the photovoltaic panel pvtheory It can be calculated by the following formula:

[0040] E pvtheory (I) = IS pv η pv η sys

[0041] Where I is the solar radiation intensity, S pv is the photovoltaic panel installation area, η pv is the photovoltaic panel conversion efficiency, η sys For the overall system efficiency;

[0042] Based on the convolutional neural network (CNN), feature vectors are selected from historical data, and W is selected. i The closest historical time point i with the same weather and the same value. For each time point i, construct the feature vector Z i , including the solar radiation intensity I i , temperature T i Weather conditions n , as well as the changes in solar radiation intensity, temperature, and weather conditions in the previous h hours, the corresponding target value is the actual photovoltaic panel power generation Q at that time point i ;

[0043] Set up the convolution layer: Set k convolution kernels, each with a dimension of f×f. The convolution kernel slides on the input feature vector to perform convolution operations to obtain the convolution result. Set the output of the convolution layer to be C out , then the convolution operation formula is:

[0044]

[0045] Where k(x,y) is the element with coordinates (x,y) of the convolution kernel, and Z i (x+i, y+j) is the element of the input feature vector at the corresponding coordinate;

[0046] A pooling layer is set after the convolution layer. The function of the pooling layer is to downsample the output of the convolution layer, reduce the amount of data and extract the main features. By setting the maximum pooling, the pooling window size is set to p×p, and the maximum pooling operation formula is:

[0047]

[0048] Among them, M out (x,y) is the element with coordinates (x,y) output by the pooling layer, C out (x,y) is the element with coordinates (x,y) output by the convolutional layer;

[0049] After being processed by multiple convolutional layers and pooling layers, it enters the fully connected layer, and maps the refined feature representation to the predicted power generation value through a fully connected layer. The weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Q pred for:

[0050] Q pred =W out ·h m +b out

[0051] Among them, h m It is a refined feature representation obtained after processing through multiple convolutional layers and pooling layers, optimized through training data, and then using the stochastic gradient descent (SGD) optimization algorithm to make the predicted value Q pred As close as possible to the actual power generation value Q i ;

[0052] Construct a hybrid prediction model for photovoltaic panel power generation, and assume that the final predicted photovoltaic panel power generation per hour on the second day is E pv (t), where t represents the hour index, ranging from 1 to 24, and the prediction results of the two models are fused by weighted average:

[0053] E pv (t) = βE pvtheory (I next (t))+(1-β)Q pred (t)

[0054] Among them, β is the weight coefficient, which ranges from 0 to 1. It is set to 0.5 in the initial stage and adjusted according to the prediction results. next (t) is the predicted solar radiation intensity at hour t of the next day, Q pred (t) is the power generation at hour t of the next day predicted by the CNN network.

[0055] S4: Based on the historical sum of wind power generation and photovoltaic panel power generation and power consumption data, as well as the forecast of future power consumption, determine the charging decision limit of the energy storage system: ensure that the energy storage capacity can have sufficient power reserves to meet the power demand when the sum of wind power generation and photovoltaic panel power generation is insufficient, and can effectively store the excess power when the sum of wind power generation and photovoltaic panel power generation is excessive;

[0056] Step S4 specifically determines the charging decision limit of the energy storage system based on the predicted value of the photovoltaic panel power generation and the predicted value of the wind turbine power generation, as well as the prediction of future power consumption:

[0057]

[0058] Among them, C storage is the rated capacity of the energy storage system, L(t) is the predicted power consumption at hour t, E pv (t is the predicted value of photovoltaic panel power generation, P w (t) is the predicted value of wind turbine power generation.

[0059] S5: Determine the charging decision function: When the sum of wind power generation and photovoltaic panel power generation is greater than power consumption and the energy storage system is not fully charged, calculate the charging power and adjust the charging power taking into account the market electricity price factor; Step S5 determines the charging decision function specifically that when the following conditions are met at the same time, the energy storage system performs charging operation, and the sum of wind power generation and photovoltaic panel power generation is greater than power consumption, that is, P w (t)+E pv (t)-L(t)>0, the energy storage system is not fully charged, that is, S(t) max In this case, the charging power P charge (t) is calculated as follows:

[0060]

[0061] min means taking the minimum value of the two values ​​in brackets, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at the tth hour, L(t) is the predicted power consumption at the tth hour, which is determined based on the historical power consumption data; η charge is the charging efficiency of the energy storage device, S​max is the maximum power allowed by the energy storage system, and S(t) is the power state of the energy storage system at the tth hour;

[0062] When the market electricity price is low and the above charging conditions are met, in order to obtain more benefits when the electricity price is high, the charging power can be appropriately increased. The specific settings are as follows: price (t) <P threshold And P w (t)+E pv (t)-L(t)>0 and S(t) max hour:

[0063]

[0064] Among them, P price (t) is the known market electricity price at hour t, P threshold It is the set electricity price threshold. When the market electricity price is lower than this threshold, the electricity price is considered to be low, and the condition of increasing the charging power to obtain more profits when the electricity price is high is met. k is the adjustment coefficient, k>1, which is determined by specific market conditions and experience.

[0065] S6: Determine the discharge decision function: When the sum of wind power generation and photovoltaic panel power generation is less than power consumption and there is power available in the energy storage system, calculate the discharge power and adjust the charging power by considering the market electricity price. Step S6 determines the discharge decision function as follows: When the following conditions are met at the same time, the energy storage system performs a discharge operation and the power generation is less than the power consumption, that is, P w (t)+E pv (t)-L(t)<0, there is power available in the energy storage system, that is, S(t)>S min , discharge power P discharge (t) is calculated as follows:

[0066]

[0067] min means taking the minimum value between the two values ​​in brackets, L(t) is the predicted power consumption in hour t, P w (t is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at hour t; η discharge is the discharge efficiency of the energy storage device, S(t) is the state of charge of the energy storage system at the tth hour, S min It is the minimum amount of electricity allowed by the set energy storage system;

[0068] When the market electricity price is high and the above discharge conditions are met, in order to obtain more benefits, the discharge power can be appropriately increased. The specific settings are as follows: price (t)>P threshold2 ​And P w (t)+E pv When (t)-L(t)<0:

[0069]

[0070] Among them, P price (t) is the known market electricity price at hour t, P threshold2 It is the set electricity price threshold. When the market electricity price is higher than this threshold, the electricity price is considered to be high, and the condition for obtaining more profits by increasing the discharge power is met. m is the adjustment coefficient, m>1, which is determined according to market conditions and experience.

[0071] When the power generation is equal to the power consumption, that is, P w (t)+E pv When (t)-L(t)=0, the energy storage system does not need to be charged or discharged. charge (t) = 0, P discharge (t)=0.

[0072] S7: Considering the charging optimization of battery life, the battery life is extended by optimizing the charging current, charging depth, and charging temperature factors; the charging optimization goal of step S7 considering the battery life is to construct a mathematical model that reflects the relationship between the battery life and the charging current, charging depth, and charging temperature factors. The relationship can be expressed as:

[0073]

[0074] Where L is the battery life, I c is the charging current, D c is the depth of charge, T c is the charging temperature, L0 is the initial life of the battery, N is the number of charges, a i is a coefficient that measures the relative importance of each factor on the battery life, b is a coefficient related to the number of charging times, reflecting the degree of influence of the increase in charging times on the battery life, and f i It is a function related to each factor, as follows:

[0075] For the function f1 related to the charging current:

[0076]

[0077] Among them, I c is the charging current, I c0 is the reference charging current, specifically the optimal charging current under normal conditions, I cl is the upper limit of the charging current. This function maps the value of the charging current to a specific interval in order to describe its impact on the battery life;

[0078] For the function f2 related to the depth of charge:

[0079]

[0080] Among them, D c is the depth of charge, D c0 It is the reference charging depth, specifically the ideal charging depth when the battery is in normal use, D c1 is the upper limit of the depth of charge. This function also maps the value of the depth of charge to a specific interval to reflect its impact on battery life;

[0081] For the function f3 related to charging temperature:

[0082]

[0083] Among them, T c is the charging temperature, T c0 is the suitable operating temperature of the battery, T c1 is the upper limit of the charging temperature. This function maps the value of the charging temperature to a specific interval to describe its impact on battery life;

[0084] Use the gradient descent optimization algorithm to gradually find the maximum value of the battery life function by updating the value of the variable along the gradient direction of the function, that is, the direction in which the function changes fastest;

[0085] Randomly change the charging current from 0 to the maximum allowed I max Initialization charging current I c0 , from 0 to the maximum allowed charging depth D max Initialize the charge depth D c0 And the lowest temperature T at which the battery can work normally min To the maximum temperature T max Initialize charging temperature T c0 , set the learning rate α, the initial value is 0.01, adjust it according to the specific situation, and calculate the partial derivatives of the objective function with respect to each variable for each iteration k;

[0086] According to the formula of the gradient descent method, update the value of the charging current:

[0087]

[0088] in To update the charging current, is the initial charging current, α is the learning rate, For L with respect to the charging current I c The partial derivative of

[0089] Likewise, update the value of the depth of charge:

[0090]

[0091] in To update the charge depth, is the initial charging depth, α is the learning rate, L is about the charge depth D c The partial derivative of

[0092] Update the value of charging temperature:

[0093]

[0094] in Update charging temperature, is the initial charging temperature, α is the learning rate, About charging current I c The partial derivative of

[0095] When the difference between the updated value of each variable and the value of the previous iteration is less than the set threshold in several consecutive iterations, the algorithm stops iterating. At this time, the current combination of charging current, charging depth and charging temperature is the approximate optimal solution for the longest battery life. By adopting the above-mentioned charging optimization scheme based on the gradient descent method, it is possible to find the optimal combination that can maximize the battery life while meeting the charging demand by optimizing the charging current, charging depth and charging temperature factors, thereby effectively extending the battery life.

[0096] S8: According to the specific situation, continuously adjust and optimize the weight coefficients in the hybrid prediction model, the adjustment coefficients in the energy storage charging and discharging strategy, the rated capacity of the energy storage system, the minimum and maximum allowable power and other related parameters to improve the system operation efficiency and economic benefits.

[0097] Compared with the prior art, the present invention has the following beneficial effects:

[0098] By collecting future forecast data from local meteorological departments and comparing it with historical data, a forecasting mathematical algorithm is established. A hybrid model is constructed by combining a variety of different types of forecasting algorithms to predict the power generation of energy generation equipment. The power generation is predicted by selecting the time point with the same weather conditions and the closest historical weather conditions. Through mathematical algorithms, the charging and discharging operations of energy storage can be planned in advance to better absorb new energy.

[0099] Combine different types of prediction algorithms to build a hybrid model. Combine traditional models based on physical principles with data-driven machine learning models. Physical models can provide a basic theoretical framework to ensure that predictions are within a reasonable range, while machine learning models can better capture nonlinear relationships and complex patterns in the data and improve the accuracy of predictions. The power generation of energy generation equipment can be predicted more accurately, thereby providing a more reliable basis for the subsequent charging and discharging operation planning of energy storage.

[0100] Based on the predicted power generation of photovoltaic panels and wind turbines, as well as the prediction of future electricity consumption, the charging decision limit of the energy storage system is determined to improve the accuracy of future charging and discharging predictions.

[0101] When determining the charging decision function and the discharging decision function, not only the relationship between power generation and power consumption and the power status of the energy storage system are taken into consideration, but the charging power and discharging power are also adjusted in combination with market electricity price factors, which can better achieve the improvement of economic benefits.

[0102] The gradient descent optimization algorithm is used to optimize the charging parameters related to battery life. By updating the values ​​of the variables along the function gradient direction, the maximum value of the battery life function is gradually found, thereby determining the optimal charging parameter combination that maximizes the battery life. This optimization scheme based on the gradient descent method can more efficiently and accurately find the optimal solution that meets the charging needs while extending the battery life, effectively extending the battery service life.

[0103] According to the specific situation, the weight coefficients in the hybrid prediction model, the adjustment coefficients in the energy storage charging and discharging strategy, the rated capacity of the energy storage system, the minimum and maximum allowable power and other related parameters are continuously adjusted and optimized. This dynamic adjustment mechanism enables the entire system to better adapt to different operating environments, market conditions and equipment status changes, further improve the system's operating efficiency and economic benefits, and maintain the system's good performance in long-term operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 This is a diagram of the economic operation control method and system steps for a distributed energy storage system for absorbing new energy sources according to the present invention; DETAILED DESCRIPTION

[0105] The technical solutions in the embodiments of the present invention will be fully described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0106] like Figure 1 As shown, a method and system for controlling the economic operation of distributed energy storage to absorb new energy is characterized by comprising the following steps:

[0107] S1: Data collection and preprocessing: Obtain the meteorological forecast data of wind power and photovoltaic power stations from the local meteorological departments, and collect the corresponding historical meteorological data of the past few years to establish the data time series of meteorological factors;

[0108] S2: Predict wind power generation: Predict wind power generation based on the physical principle model and the long short-term memory network LSTM data-driven machine learning model respectively, and build a hybrid prediction model to fuse the prediction results of the physical principle model and the LSTM model in a weighted average manner to obtain the final wind turbine hourly power generation prediction value;

[0109] S3: Predict the power generation of photovoltaic panels: Predict the power generation of photovoltaic panels based on the data-driven machine learning model of the physical principle model and the convolutional neural network (CNN) and build a hybrid prediction model to calculate the final hourly power generation prediction value of the photovoltaic panel by weighted average fusion of the prediction results of the physical principle model and the CNN model;

[0110] S4: Based on the historical sum of wind power generation and photovoltaic panel power generation and power consumption data, as well as the forecast of future power consumption, determine the charging decision limit of the energy storage system to ensure that the energy storage capacity can have sufficient power reserves to meet the power demand when the sum of wind power generation and photovoltaic panel power generation is insufficient, and can effectively store the excess power when the sum of wind power generation and photovoltaic panel power generation is excessive;

[0111] S5: Determine the charging decision function: When the sum of wind power generation and photovoltaic panel power generation meets the power consumption and the energy storage system is not fully charged, calculate the charging power and adjust the charging power considering the market electricity price factor;

[0112] S6: Determine the discharge decision function: When the sum of wind power generation and photovoltaic panel power generation is less than the power consumption and there is power available in the energy storage system, calculate the discharge power and adjust the charging power by considering the market electricity price factor;

[0113] S7: Optimize charging current, charging depth, and charging temperature to optimize battery life, and extend the battery life as much as possible while meeting charging requirements.

[0114] S8: According to the specific situation, the weight coefficients in the hybrid prediction model, the adjustment coefficients in the charging and discharging decision functions in the above steps, and the rated capacity of the energy storage system are continuously adjusted and optimized to improve the system operation efficiency and economic benefits.

[0115] Step S1 obtains the weather forecast data of wind power and photovoltaic power from the local meteorological department, collects the corresponding historical weather data of the past few years, and establishes the time series data of various meteorological factors: including wind speed v next、Solar radiation intensity I next , Temperature T next 、Weather conditions next , and collect historical meteorological data from the past few years, which are recorded as wind speed series {v i}、Solar irradiance intensity sequence {I i}、Temperature series {T i}、Weather condition sequence {W n}, where i represents the time index.

[0116] Weather conditions next It is determined by assigning values ​​to the specific weather conditions, specifically, the clear value is 1, cloudy is 2, overcast is 3, light rain is 4, moderate rain and above is 5, snow is 6, fog is 7, strong wind is 8, and thunderstorm is 9.

[0117] Step S2 predicts the power generation of the wind turbine generator according to the physical principle model, that is, the theoretical power output P of the wind turbine generator. wtheory (v) is calculated by the following formula:

[0118]

[0119] Where ρ is the air density, A is the swept area of ​​the wind turbine blade, C p is the wind energy utilization coefficient, v is the wind speed;

[0120] The LSTM data-driven machine learning model selects feature vectors from historical data and selects W i The closest historical time point i with the same weather and the same value. For each time point i, the eigenvector X i Including wind speed v i , Temperature T i 、Weather conditions n and the changes in wind speed, temperature, and weather conditions in the previous h hours; the corresponding target value is the actual wind turbine power generation Y at that time point i ;

[0121] Assume that the input dimension of the LSTM network is d, that is, the feature vector X i The dimension of the hidden layer is m, the output dimension is 1, and the LSTM unit formula is as follows:

[0122] Forget Gate:

[0123] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0124] Among them, f t is the output of the forget gate; σ is the sigmoid function, which maps the output value to between 0 and 1, W f is the weight matrix of the forget gate, with a dimension of (m+d)×m, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous moment, with dimension m, X t is the input feature vector at the current moment, with dimension d;

[0125] Set up the input gate:

[0126] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0127]

[0128] Among them, i t is the output of the input gate, is the new candidate cell state, W i , W C are the corresponding weight matrices with dimensions of (m+d)×m and (m+d)×m, respectively, b i , b C is the corresponding bias vector;

[0129] Set the cell status update:

[0130]

[0131] Among them, C t is the updated cell state, (f t ·C t-1 ) represents the old information that the forget gate decides to forget and represents the new information that the input gate decides to update, C t-1 is the cell state at the previous moment;

[0132] Set up the output gate:

[0133] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0134] h t =o t tanh(C t )

[0135] Among them,t is the output of the output gate, h t is the hidden state at the current moment, W o is the weight matrix of the output gate, with dimension (m+d)×m, b o is the bias vector of the output gate;

[0136] After setting to N time steps, N corresponds to the length of the time series in the historical data, the hidden state h is transformed through a fully connected layer. N Mapped to the predicted power generation value, the weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Y pred for:

[0137] Y pred =W out ·h N +b out

[0138] where Y pred Represents the predicted power generation value, W out is the weight matrix of the fully connected layer, W out This is the weight matrix of the fully connected layer, b out is the bias vector of the fully connected layer;

[0139] Through training data, that is, feature vector X i and target value Y i For all weight matrices (W f , W i , W C , W o , W out ) and the bias vector (b f 、b i 、b C 、b o 、b o 、b out ) is optimized using the stochastic gradient descent SGD optimization algorithm so that the predicted value Y pred As close as possible to the actual power generation value Y i ;

[0140] Combining the above physical principle model with the data-driven machine learning model, a hybrid prediction model is constructed. The final predicted hourly power generation of the wind turbine on the second day is P w (t), the prediction results of the two models are fused by weighted average:

[0141] P w (t) = αP wtheory (v next(t))+(1-α)Y pred (t)

[0142] Among them, t represents the hour index, ranging from 1 to 24, α is the weight coefficient, ranging from 0 to 1, and is set to 0.5 in the initial stage and adjusted according to the prediction effect, v next (t) is the predicted wind speed at hour t of the next day, Y pred (t) is the power generation at the t hour of the next day predicted by the LSTM network.

[0143] Step S3 predicts the power generation of the photovoltaic panel based on the physical principle model, that is, the power generation E of the photovoltaic panel pvtheory Calculated by the following formula:

[0144] E pvtheory (I) = IS pv η pv η sys

[0145] Where I is the solar radiation intensity, S pv is the photovoltaic panel installation area, η pv is the photovoltaic panel conversion efficiency, η sys For the overall system efficiency;

[0146] Based on the convolutional neural network CNN, feature vectors are selected from historical data, and W is selected i The closest historical time point i with the same weather and the same value. For each time point i, construct the feature vector Z i , including the solar radiation intensity I i , temperature T i Weather conditions n , as well as the changes in solar radiation intensity, temperature, and weather conditions in the previous h hours, the corresponding target value is the actual photovoltaic panel power generation Q at that time point i ;

[0147] Set up the convolution layer: Set k convolution kernels, each with a dimension of f×f. The convolution kernel slides on the input feature vector to perform convolution operations to obtain the convolution result. Set the output of the convolution layer to be C out , then the convolution operation formula is:

[0148]

[0149] Where k(x,y) is the element with coordinates (x,y) of the convolution kernel, and Z i (x+i, y+j) is the element of the input feature vector at the corresponding coordinate;

[0150] A pooling layer is set after the convolution layer. The function of the pooling layer is to downsample the output of the convolution layer, reduce the amount of data and extract the main features. By setting the maximum pooling, the pooling window size is set to p×p, and the maximum pooling operation formula is:

[0151]

[0152] Among them, M out (x,y) is the element with coordinates (x,y) output by the pooling layer, C out (x,y) is the element with coordinates (x,y) output by the convolutional layer;

[0153] After being processed by multiple convolutional layers and pooling layers, it enters the fully connected layer, and maps the refined feature representation to the predicted power generation value through a fully connected layer. The weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Q pred for:

[0154] Q pred =W out ·h m +b out

[0155] Among them, h m It is a refined feature representation obtained after processing through multiple convolutional layers and pooling layers, optimized through training data, and then using the stochastic gradient descent SGD optimization algorithm to make the predicted value Q pred As close as possible to the actual power generation value Q i ;

[0156] Construct a hybrid prediction model for photovoltaic panel power generation, and assume that the final predicted photovoltaic panel power generation per hour on the second day is E pv (t), where t represents the hour index, ranging from 1 to 24, and the prediction results of the two models are fused by weighted average:

[0157] E pv (t) = βE pvtheory (I next (t))+(1-β)Q pred (t)

[0158] Among them, β is the weight coefficient, which ranges from 0 to 1. It is set to 0.5 in the initial stage and adjusted according to the prediction results. next (t) is the predicted solar radiation intensity at hour t of the next day, Q pred (t) is the power generation at hour t of the next day predicted by the CNN network.

[0159] Step S4 specifically determines the charging decision limit of the energy storage system based on the sum of the historical wind power generation and photovoltaic panel power generation and power consumption data, as well as the prediction of future power consumption:

[0160]

[0161] Among them, C storage is the rated capacity of the energy storage system, L(t) is the predicted power consumption at hour t, E pv (t) is the predicted value of photovoltaic panel power generation, P w (t) is the predicted value of wind turbine power generation.

[0162] Step S5 determines the charging decision function. Specifically, when the following conditions are met at the same time, the energy storage system performs charging operation, and the sum of wind power generation and photovoltaic panel power generation is greater than power consumption, that is, P w (t)+E pv (t)-L(t)>0, the energy storage system is not fully charged, that is, S(t) max In this case, the charging power P charge (t) is calculated as follows:

[0163]

[0164] min means taking the minimum value of the two values ​​in brackets, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at the tth hour, L(t) is the predicted power consumption at the tth hour, which is determined based on the historical power consumption data; η charge is the charging efficiency of the energy storage device, S max is the maximum power allowed by the energy storage system, and S(t) is the power state of the energy storage system at the tth hour;

[0165] When the market electricity price is low and the above charging conditions are met, in order to obtain more benefits when the electricity price is high, the charging power is appropriately increased. The specific settings are as follows: price (t) <P threshold And P w (t)+E pv (t)-L(t)>0 and S(t) max hour:

[0166]

[0167] Among them, P price (t) is the known market electricity price at hour t, P threshold ​​It is the set electricity price threshold. When the market electricity price is lower than this threshold, the electricity price is considered to be low, and the condition of increasing the charging power to obtain more profits when the electricity price is high is met. k is the adjustment coefficient, k>1, which is determined by specific market conditions and experience.

[0168] Step S6 determines the discharge decision function. Specifically, when the following conditions are met at the same time, the energy storage system performs a discharge operation, and the sum of the wind power generation and the photovoltaic panel power generation is less than the power consumption, that is, P w (t)+E pv (t)-L(t)<0, there is power available in the energy storage system, that is, S(t)>S min , discharge power P discharge (t) is calculated as follows:

[0169]

[0170] min means taking the minimum value, L(t) is the predicted power consumption in the tth hour, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at hour t; η discharge is the discharge efficiency of the energy storage device, S(t) is the state of charge of the energy storage system at the tth hour, S min It is the minimum amount of electricity allowed by the set energy storage system;

[0171] When the market electricity price is high and the above discharge conditions are met, in order to obtain more benefits, the discharge power is appropriately increased. The specific settings are as follows: price (t)>P threshold2 And P w (t)+E pv When (t)-L(t)<0:

[0172]

[0173] Among them, P price (t) is the known market electricity price at hour t, P threshold2 It is the set electricity price threshold. When the market electricity price is higher than this threshold, the electricity price is considered to be high, and the condition for obtaining more profits by increasing the discharge power is met. m is the adjustment coefficient, m>1, which is determined according to market conditions and experience.

[0174] When the sum of wind power generation and photovoltaic panel power generation is equal to the power consumption, that is, P w (t)+E pv When (t)-L(t)=0, the energy storage system does not need to be charged or discharged. charge (t) = 0, P discharge (t)=0.

[0175] 10. The method for controlling the economic operation of a distributed energy storage system for absorbing new energy according to claim 1 is characterized in that the charging optimization target of considering the battery life in step S7 is specifically to construct a mathematical model reflecting the relationship between the battery life and the charging current, charging depth and charging temperature factors, and the relationship can be expressed as:

[0176]

[0177] Where L is the battery life, I c is the charging current, D c is the depth of charge, T c is the charging temperature, L0 is the initial life of the battery, N is the number of charges, a i It is a coefficient that measures the relative importance of each factor on the battery life, and b is a coefficient related to the number of charging times, reflecting the degree of influence of the increase in charging times on the battery life;

[0178] f i (I c ,D c ,T c ) are calculated separately, which specifically includes f1(I c )、f2(D c )、f3(T c ), as follows:

[0179] For the function f1(I c ):

[0180]

[0181] Among them, I c is the charging current, I c0 is the reference charging current, specifically the optimal charging current under normal conditions, I cl is the upper limit of the charging current. This function maps the value of the charging current to a specific interval in order to describe its impact on the battery life;

[0182] For the function f2(D c ):

[0183]

[0184] Among them, D c is the depth of charge, D c0 It is the reference charging depth, specifically the ideal charging depth when the battery is in normal use, D c1 is the upper limit of the depth of charge. This function also maps the value of the depth of charge to a specific interval to reflect its impact on battery life;

[0185] For the function f3(T c ):

[0186]

[0187] Among them, T c is the charging temperature, T c0 is the suitable operating temperature of the battery, T c1 is the upper limit of the charging temperature. This function maps the value of the charging temperature to a specific interval to describe its impact on battery life;

[0188] Use the gradient descent optimization algorithm to gradually find the maximum value of the battery life function by updating the value of the variable along the gradient direction of the function, that is, the direction in which the function changes fastest;

[0189] Randomly change the charging current from 0 to the maximum allowed I max Initialize charging current I c0 , from 0 to the maximum allowed charging depth D max Initialize the charge depth D c0 And the lowest temperature T at which the battery can work normally min To the maximum temperature T max Initialize charging temperature T c0 , set the learning rate α, the initial value is 0.01, adjust it according to the specific situation, and calculate the partial derivatives of the objective function with respect to each variable for each iteration k;

[0190] According to the formula of the gradient descent method, update the value of the charging current:

[0191]

[0192] in To update the charging current, is the initial charging current, α is the learning rate, For L with respect to the charging current I c The partial derivative of

[0193] Likewise, update the value of the depth of charge:

[0194]

[0195] in To update the charge depth, is the initial charging depth, α is the learning rate, L is about the charge depth D c The partial derivative of

[0196] Update the value of charging temperature:

[0197]

[0198] in Update charging temperature, is the initial charging temperature, α is the learning rate, About charging current I c The partial derivative of

[0199] When the difference between the updated value of each variable and the value of the previous iteration is less than the set threshold in several consecutive iterations, the algorithm stops iterating. At this time, the current combination of charging current, charging depth and charging temperature is the approximate optimal solution for the longest battery life.

[0200] The server stores a system program, wherein the processor implements the steps of the method according to any one of claims 1 to 9 when executing the computer program.

[0201] The following is the implementation method. The weather forecast data of wind power and photovoltaic power stations are obtained from the local meteorological department, and the corresponding historical meteorological data of the past few years are collected to establish the time series data of various meteorological factors; specifically, the wind speed v next 、Solar radiation intensity I next , Temperature T next 、Weather conditions next , and collect historical meteorological data from the past few years, which are recorded as wind speed series {v i}、Solar irradiance intensity sequence {I i}、Temperature series {T i}、Weather condition sequence {W n}, where i represents the time index; weather conditions W next It is determined by assigning values ​​to the specific weather conditions, specifically, the clear value is 1, cloudy is 2, overcast is 3, light rain is 4, moderate rain and above is 5, snow is 6, fog is 7, strong wind is 8, and thunderstorm is 9.

[0202] For the prediction of wind turbine power generation, the theoretical power output is calculated based on the physical principle model, the air density, the swept area of ​​the impeller, and the wind energy utilization coefficient according to the aerodynamic principle. The feature vector containing wind speed, temperature, weather conditions and their changes is selected by using the long short-term memory network (LSTM) data-driven machine learning model. The power generation is predicted through the LSTM unit-related operations and the fully connected layer mapping steps, and the weight matrix and bias vector are optimized through training. A hybrid prediction model is constructed, and the prediction results of the physical principle model and the LSTM model are integrated by weighted average to obtain the final wind turbine hourly power generation prediction value.

[0203] The prediction of wind turbine power generation is based on the aerodynamic principle of wind power generation. The theoretical power output P of the wind turbine is wtheory (v) is calculated by the following formula:

[0204]

[0205] Where ρ is the air density, A is the swept area of ​​the wind turbine blade, C p is the wind energy utilization coefficient;

[0206] The machine learning model driven by the long short-term memory network (LSTM) selects feature vectors from historical data and selects W i The closest historical time point i with the same weather and the same value. For each time point i, the eigenvector X i Including wind speed v i , Temperature T i 、Weather conditions n and the changes in wind speed, temperature, and weather conditions in the previous h hours; the corresponding target value is the actual wind turbine power generation Y at that time point i ;

[0207] Assume that the input dimension of the LSTM network is d, that is, the feature vector X i The dimension of the hidden layer is m, the output dimension is 1, and the LSTM unit formula is as follows:

[0208] Forget Gate:

[0209] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0210] Among them, f t is the output of the forget gate; σ is the sigmoid function, which maps the output value to between 0 and 1, W f is the weight matrix of the forget gate, with dimension (m+d)×m, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous moment, with dimension m, X t is the input feature vector at the current moment, with dimension d;

[0211] Set up the input gate:

[0212] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0213]

[0214] Among them, i t is the output of the input gate, is the new candidate cell state, W i , W C are the corresponding weight matrices with dimensions of (m+d)×m and (m+d)×m, respectively, b i , b C is the corresponding bias vector;

[0215] Set the cell status update:

[0216]

[0217] Among them, C t is the updated cell state, (f t ·C t-1 ) represents the old information that the forget gate decides to forget and represents the new information that the input gate decides to update, C t-1 is the cell state at the previous moment;

[0218] Set up the output gate:

[0219] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0220] h t =o t tanh(C t )

[0221] Among them, t is the output of the output gate, h t is the hidden state at the current moment, W o is the weight matrix of the output gate, with dimension (m+d)×m, b o is the bias vector of the output gate;

[0222] After setting to N time steps, N corresponds to the length of the time series in the historical data, the hidden state h is transformed through a fully connected layer. N Mapped to the predicted power generation value, the weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Y pred for:

[0223] Y pred =W out ·h N +bout

[0224] where Y pred Represents the predicted power generation value, W out is the weight matrix of the fully connected layer, W out This is the weight matrix of the fully connected layer, b out is the bias vector of the fully connected layer;

[0225] Through training data, that is, feature vector X i and target value Y i For all weight matrices (W f , W i , W C , W o , W out ) and the bias vector (b f , b i , b C , b o , b o , b out ) is optimized using the stochastic gradient descent (SGD) optimization algorithm so that the predicted value Y pred As close as possible to the actual power generation value Y i ;

[0226] Combining the above two models, a hybrid prediction model is constructed. The final predicted hourly power generation of the wind turbine on the second day is P w (t), the prediction results of the two models are fused by weighted average:

[0227] P w (t) = αP wtheory (v next (t))+(1-α)Y pred (t)

[0228] Among them, t represents the hour index, ranging from 1 to 24, α is the weight coefficient, ranging from 0 to 1, and is set to 0.5 in the initial stage and adjusted according to the prediction effect, v next (t) is the predicted wind speed at hour t of the next day, Y pred (t) is the power generation at the t hour of the next day predicted by the LSTM network.

[0229] For the prediction of photovoltaic panel power generation, the physical principle model is used to calculate the theoretical power generation according to the photovoltaic power generation principle combined with the solar radiation intensity, installation area, conversion efficiency, and system comprehensive efficiency parameters. The data-driven machine learning model based on the convolutional neural network (CNN) is used to select the characteristic vectors containing the solar radiation intensity, temperature, weather conditions and their changes, and the power generation is predicted through the convolution layer, pooling layer processing and fully connected layer mapping operations, and the relevant parameters are trained and optimized. A hybrid prediction model is constructed to fuse the prediction results of the physical principle model and the CNN model by weighted average to calculate the final predicted value of the photovoltaic panel power generation per hour.

[0230] The prediction of photovoltaic panel power generation is based on the principle of photovoltaic power generation. pvtheory It can be calculated by the following formula:

[0231] E pvtheory (I) = IS pv η pv η sys

[0232] Where I is the solar radiation intensity, S pv is the photovoltaic panel installation area, η pv is the photovoltaic panel conversion efficiency, η sys For the overall system efficiency;

[0233] Based on the convolutional neural network (CNN), feature vectors are selected from historical data, and W is selected. i The closest historical time point i with the same weather and the same value. For each time point i, construct the feature vector Z i , including the solar radiation intensity I i , temperature T i Weather conditions n , as well as the changes in solar radiation intensity, temperature, and weather conditions in the previous h hours, the corresponding target value is the actual photovoltaic panel power generation Q at that time point i ;

[0234] Set up the convolution layer: Set k convolution kernels, each with a dimension of f×f. The convolution kernel slides on the input feature vector to perform convolution operations to obtain the convolution result. Set the output of the convolution layer to be C out , then the convolution operation formula is:

[0235]

[0236] Where k(x,y) is the element with coordinates (x,y) of the convolution kernel, and Z i (x+i, y+j) is the element of the input feature vector at the corresponding coordinate;

[0237] A pooling layer is set after the convolution layer. The function of the pooling layer is to downsample the output of the convolution layer, reduce the amount of data and extract the main features. By setting the maximum pooling, the pooling window size is set to p×p, and the maximum pooling operation formula is:

[0238]

[0239] Among them, M out (x,y) is the element with coordinates (x,y) output by the pooling layer, C out (x,y) is the element with coordinates (x,y) output by the convolutional layer;

[0240] After being processed by multiple convolutional layers and pooling layers, it enters the fully connected layer, and maps the refined feature representation to the predicted power generation value through a fully connected layer. The weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Q pred for:

[0241] Q pred =W out ·h m +b out

[0242] Among them, h m It is a refined feature representation obtained after processing through multiple convolutional layers and pooling layers, optimized through training data, and then using the stochastic gradient descent (SGD) optimization algorithm to make the predicted value Q pred As close as possible to the actual power generation value Q i ;

[0243] Construct a hybrid prediction model for photovoltaic panel power generation, and assume that the final predicted photovoltaic panel power generation per hour on the second day is E pv (t), where t represents the hour index, ranging from 1 to 24, and the prediction results of the two models are fused by weighted average:

[0244] E pv (t) = βE pvtheory (I next (t))+(1-β)Q pred (t)

[0245] Among them, β is the weight coefficient, which ranges from 0 to 1. It is set to 0.5 in the initial stage and adjusted according to the prediction results. next (t) is the predicted solar radiation intensity at hour t of the next day, Q pred (t) is the power generation at hour t of the next day predicted by the CNN network.

[0246] S4: Determine the charging decision limit of the energy storage system based on historical power generation and power consumption data, as well as the forecast of future power consumption. The core idea is to ensure that the energy storage capacity can have sufficient power reserves to meet power demand when power generation is insufficient, and to effectively store excess power when power generation is excessive.

[0247] Based on the predicted power generation of the photovoltaic panels and the wind turbines, as well as the prediction of future power consumption, the charging decision limit of the energy storage system is determined:

[0248]

[0249] Among them, C storage is the rated capacity of the energy storage system, L(t) is the predicted power consumption at hour t, E pv (t is the predicted value of photovoltaic panel power generation, P w (t) is the predicted value of wind turbine power generation.

[0250] Determine the charging decision function. When the power generation is greater than the power consumption and the energy storage system is not fully charged, calculate the charging power and adjust the charging power by considering the market electricity price. Determine the charging decision function specifically that when the following conditions are met at the same time, the energy storage system performs charging operation and the power generation is greater than the power consumption, that is, P w (t)+E pv (t)-L(t)>0, the energy storage system is not fully charged, that is, S(t) max In this case, the charging power P charge (t) is calculated as follows:

[0251]

[0252] min means taking the minimum value of the two values ​​in brackets, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at the tth hour, L(t) is the predicted power consumption at the tth hour, which is determined based on the historical power consumption data; η charge is the charging efficiency of the energy storage device, S max is the maximum power allowed by the energy storage system, and S(t) is the power state of the energy storage system at the tth hour;

[0253] When the market electricity price is low and the above charging conditions are met, in order to obtain more benefits when the electricity price is high, the charging power can be appropriately increased. The specific settings are as follows: price (t) <P threshold And P w (t)+E pv (t)-L(t)>0 and S(t) <S​max hour:

[0254]

[0255] Among them, P price (t) is the known market electricity price at hour t, P threshold It is the set electricity price threshold. When the market electricity price is lower than this threshold, the electricity price is considered to be low, and the condition of increasing the charging power to obtain more profits when the electricity price is high is met. k is the adjustment coefficient, k>1, which is determined by specific market conditions and experience.

[0256] Determine the discharge decision function. When the power generation is less than the power consumption and there is power available in the energy storage system, calculate the discharge power and adjust the charging power by considering the market electricity price. Determine the discharge decision function as follows: When the following conditions are met at the same time, the energy storage system performs the discharge operation and the power generation is less than the power consumption, that is, P w (t)+E pv (t)-L(t)<0, there is power available in the energy storage system, that is, S(t)>S min , discharge power P discharge (t) is calculated as follows:

[0257]

[0258] min means taking the minimum value between the two values ​​in brackets, L(t) is the predicted power consumption in hour t, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at hour t; η discharge is the discharge efficiency of the energy storage device, S(t) is the state of charge of the energy storage system at the tth hour, S min It is the minimum amount of electricity allowed by the set energy storage system;

[0259] When the market electricity price is high and the above discharge conditions are met, in order to obtain more benefits, the discharge power can be appropriately increased. The specific settings are as follows: price (t)>P threshold2 And P w (t)+E pv When (t)-L(t)<0:

[0260]

[0261] Among them, P price (t) is the known market electricity price at hour t, P threshold2It is the set electricity price threshold. When the market electricity price is higher than this threshold, the electricity price is considered to be high, and the condition for obtaining more profits by increasing the discharge power is met. m is the adjustment coefficient, m>1, which is determined according to market conditions and experience.

[0262] When the power generation is equal to the power consumption, that is, P w (t)+E pv When (t)-L(t)=0, the energy storage system does not need to be charged or discharged. charge (t) = 0, P discharge (t)=0.

[0263] The goal of charging optimization considering battery life is to meet charging requirements and extend the battery life as much as possible, which is achieved by optimizing charging current, charging depth, and charging temperature factors. The specific goal of charging optimization considering battery life is to build a mathematical model that reflects the relationship between battery life and charging current, charging depth, and charging temperature factors. The relationship can be expressed as:

[0264]

[0265] Where L is the battery life, I c is the charging current, D c is the depth of charge, T c is the charging temperature, L0 is the initial life of the battery, N is the number of charges, a i It is a coefficient that measures the relative importance of each factor on the battery life, and b is a coefficient related to the number of charging times, reflecting the degree of influence of the increase in charging times on the battery life;

[0266] For f in the above formula i (I c ,D c ,T c ) are calculated separately, which specifically includes f1(I c )、f2(D c )、f3(T c ), as follows: For the function f1(I c ):

[0267]

[0268] Among them, I c is the charging current, I c0 is the reference charging current, specifically the optimal charging current under normal conditions, I cl is the upper limit of the charging current. This function maps the value of the charging current to a specific interval in order to describe its impact on the battery life;

[0269] For the function f2(Dc ):

[0270]

[0271] Among them, D c is the depth of charge, D c0 It is the reference charging depth, specifically the ideal charging depth when the battery is in normal use, D c1 is the upper limit of the depth of charge. This function also maps the value of the depth of charge to a specific interval to reflect its impact on battery life;

[0272] For the function f3(T c ):

[0273]

[0274] Among them, T c is the charging temperature, T c0 is the suitable operating temperature of the battery, T c1 is the upper limit of the charging temperature. This function maps the value of the charging temperature to a specific interval to describe its impact on battery life;

[0275] Use the gradient descent optimization algorithm to gradually find the maximum value of the battery life function by updating the value of the variable along the gradient direction of the function, that is, the direction in which the function changes fastest;

[0276] Randomly change the charging current from 0 to the maximum allowed I max Initialize charging current I c0 , from 0 to the maximum allowed charging depth D max Initialize the charge depth D c0 And the lowest temperature T at which the battery can work normally min To the maximum temperature T max Initialize charging temperature T c0 , set the learning rate α, the initial value is 0.01, adjust it according to the specific situation, and calculate the partial derivatives of the objective function with respect to each variable for each iteration k;

[0277] According to the formula of the gradient descent method, update the value of the charging current:

[0278]

[0279] in To update the charging current, is the initial charging current, α is the learning rate, For L with respect to the charging current I c The partial derivative of

[0280] Likewise, update the value of the depth of charge:

[0281]

[0282] in To update the charge depth, is the initial charging depth, α is the learning rate, L is about the charge depth D c The partial derivative of

[0283] Update the value of charging temperature:

[0284]

[0285] in Update charging temperature, is the initial charging temperature, α is the learning rate, About charging current I c The partial derivative of

[0286] When the difference between the updated value of each variable and the value of the previous iteration is less than the set threshold in several consecutive iterations, the algorithm stops iterating. At this time, the current combination of charging current, charging depth and charging temperature is the approximate optimal solution for the longest battery life. By adopting the above-mentioned charging optimization scheme based on the gradient descent method, it is possible to find the optimal combination that can maximize the battery life while meeting the charging demand by optimizing the charging current, charging depth and charging temperature factors, thereby effectively extending the battery life.

[0287] According to the specific situation, the weight coefficients in the hybrid prediction model, the adjustment coefficients in the energy storage charging and discharging strategy, the rated capacity of the energy storage system, the minimum and maximum allowable power and other related parameters are continuously adjusted and optimized to improve the system operation efficiency and economic benefits.

Claims

1. A method for controlling the economic operation of distributed energy storage to absorb new energy, characterized in that: The following steps are involved: S1: Data collection and preprocessing: Obtain the meteorological forecast data of wind power and photovoltaic power stations from the local meteorological departments, and collect the corresponding historical meteorological data of the past few years to establish the data time series of meteorological factors; S2: Predict wind power generation: Predict wind power generation based on the physical principle model and the long short-term memory network LSTM data-driven machine learning model respectively, and build a hybrid prediction model to fuse the prediction results of the physical principle model and the LSTM model in a weighted average manner to obtain the final wind turbine hourly power generation prediction value; S3: Predict the power generation of photovoltaic panels: Predict the power generation of photovoltaic panels based on the data-driven machine learning model of the physical principle model and the convolutional neural network (CNN) and build a hybrid prediction model to calculate the final hourly power generation prediction value of the photovoltaic panel by weighted average fusion of the prediction results of the physical principle model and the CNN model; S4: Based on the historical sum of wind power generation and photovoltaic panel power generation and power consumption data, as well as the forecast of future power consumption, determine the charging decision limit of the energy storage system to ensure that the energy storage capacity can have sufficient power reserves to meet the power demand when the sum of wind power generation and photovoltaic panel power generation is insufficient, and can effectively store the excess power when the sum of wind power generation and photovoltaic panel power generation is excessive; S5: Determine the charging decision function: When the sum of wind power generation and photovoltaic panel power generation meets the power consumption and the energy storage system is not fully charged, calculate the charging power and adjust the charging power considering the market electricity price factor; S6: Determine the discharge decision function: When the sum of wind power generation and photovoltaic panel power generation is less than the power consumption and there is power available in the energy storage system, calculate the discharge power and adjust the charging power by considering the market electricity price factor; S7: Optimize charging current, charging depth, and charging temperature to optimize battery life, and extend the battery life as much as possible while meeting charging requirements. S8: Continuously adjust and optimize the weight coefficients in the hybrid prediction model, the adjustment coefficients in the charge and discharge decision function in the above steps, and the rated capacity of the energy storage system according to the specific situation to improve the system operation efficiency and economic benefits.

2. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1 is characterized in that: Step S1 obtains the weather forecast data of wind power and photovoltaic power from the local meteorological department, collects the corresponding historical weather data of the past few years, and establishes the time series data of various meteorological factors: including wind speed v next 、Solar radiation intensity I next , Temperature T next 、Weather conditions next , and collect historical meteorological data from the past few years, which are recorded as wind speed series {v i }、Solar irradiance intensity sequence {I i }、Temperature series {T i }、Weather condition sequence {W n }, where i represents the time index.

3. The economic operation control method for accommodating new energy by distributed energy storage according to claim 2 is characterized in that: Weather conditions next It is determined by assigning values ​​to the specific weather conditions, specifically, the clear value is 1, cloudy is 2, overcast is 3, light rain is 4, moderate rain and above is 5, snow is 6, fog is 7, strong wind is 8, and thunderstorm is 9.

4. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1 is characterized in that: Step S2 predicts the power generation of the wind turbine generator as follows: Based on the physical principle model, the theoretical power output P of the wind turbine generator wtheory (v(is calculated by the following formula: Where ρ is the air density, A is the swept area of ​​the wind turbine blade, C p is the wind energy utilization coefficient, v is the wind speed; The LSTM data-driven machine learning model selects feature vectors from historical data and selects W i The closest historical time point i with the same weather and the same value. For each time point i, the eigenvector X i Including wind speed v i , Temperature T i 、Weather conditions n and the changes in wind speed, temperature, and weather conditions in the previous h hours; the corresponding target value is the actual wind turbine power generation Y at that time point i ; Assume that the input dimension of the LSTM network is d, that is, the feature vector X i The dimension of the hidden layer is m, the output dimension is 1, and the LSTM unit formula is as follows: Forget Gate: f t =σ(W f ·[h t-1 ,X t ]+b f ) Among them, f t is the output of the forget gate; σ is the sigmoid function, which maps the output value to between 0 and 1, W f is the weight matrix of the forget gate, with a dimension of (m+d)×m, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous moment, with dimension m, X t is the input feature vector at the current moment, with dimension d; Set up the input gate: i t =σ(W i ·[h t-1 ,X t ]+b i ) Among them, i t is the output of the input gate, is the new candidate cell state, W i , W C are the corresponding weight matrices with dimensions of (m+d)×m and (m+d)×m, respectively, b i , b C is the corresponding bias vector; Set the cell status update: Among them, C t is the updated cell state, (f t ·C t-1 ) represents the old information that the forget gate decides to forget and represents the new information that the input gate decides to update, C t-1 is the cell state at the previous moment; Set up the output gate: the t =σ(W o ·[h t-1 ,X t ]+b o ) h t =o t ·tanh(C t ) Among them, t is the output of the output gate, h t is the hidden state at the current moment, W o is the weight matrix of the output gate, with dimension (m+d)×m, b o is the bias vector of the output gate; After setting to N time steps, N corresponds to the length of the time series in the historical data, the hidden state h is transformed through a fully connected layer. N Mapped to the predicted power generation value, the weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Y pred for: Y pred =W out ·h N +b out where Y pred Represents the predicted power generation value, W out is the weight matrix of the fully connected layer, W out This is the weight matrix of the fully connected layer, b out is the bias vector of the fully connected layer; Through training data, that is, feature vector X i and target value Y i For all weight matrices (W f , W i , W C , W o , W out ) and the bias vector (b f 、b i 、b C 、b o 、b o 、b out ) is optimized using the stochastic gradient descent SGD optimization algorithm so that the predicted value Y pred As close as possible to the actual power generation value Y i ; Combining the above physical principle model with the data-driven machine learning model, a hybrid prediction model is constructed. The final predicted hourly power generation of the wind turbine on the second day is P w (t), the prediction results of the two models are fused by weighted average: P w (t)=αP wtheory (v next (t))+(1-α)Y pred (t) Among them, t represents the hour index, ranging from 1 to 24, α is the weight coefficient, ranging from 0 to 1, and is set to 0.5 in the initial stage and adjusted according to the prediction effect, v next (t) is the predicted wind speed at hour t of the next day, Y pred (t) is the power generation at the t hour of the next day predicted by the LSTM network.

5. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1 is characterized in that: Step S3 photovoltaic panel power generation prediction is as follows: according to the physical principle model, the photovoltaic panel power generation E pvtheory Calculated by the following formula: E pvtheory (I)=IS pv η pv η sys Where I is the solar radiation intensity, S pv is the photovoltaic panel installation area, η pv is the photovoltaic panel conversion efficiency, η sys For the overall system efficiency; Based on the convolutional neural network CNN, feature vectors are selected from historical data, and W is selected i The closest historical time point i with the same weather and the same value. For each time point i, construct the feature vector Z i , including the solar radiation intensity I i , temperature T i Weather conditions n , as well as the changes in solar radiation intensity, temperature, and weather conditions in the previous h hours, the corresponding target value is the actual photovoltaic panel power generation Q at that time point i ; Set up the convolution layer: Set k convolution kernels, each with a dimension of f×f. The convolution kernel slides on the input feature vector to perform convolution operations to obtain the convolution result. Set the output of the convolution layer to be C out , then the convolution operation formula is: Where k(x,y) is the element with coordinates (x,y) of the convolution kernel, and Z i (x+i, y+j) is the element of the input feature vector at the corresponding coordinate; A pooling layer is set after the convolution layer. The function of the pooling layer is to downsample the output of the convolution layer, reduce the amount of data and extract the main features. By setting the maximum pooling, the pooling window size is set to p×p, and the maximum pooling operation formula is: Among them, M out (x,y) is the element with coordinates (x,y) output by the pooling layer, C out (x,y) is the element with coordinates (x,y) output by the convolutional layer; After being processed by multiple convolutional layers and pooling layers, it enters the fully connected layer, and maps the refined feature representation to the predicted power generation value through a fully connected layer. The weight matrix of the fully connected layer is set to W out , the bias vector is set to b out , then the predicted power generation Q pred for: Q pred =W out ·h m +b out Among them, h m It is a refined feature representation obtained after processing through multiple convolutional layers and pooling layers, optimized through training data, and then using the stochastic gradient descent SGD optimization algorithm to make the predicted value Q pred As close as possible to the actual power generation value Q i ; Construct a hybrid prediction model for photovoltaic panel power generation, and assume that the final predicted photovoltaic panel power generation per hour on the second day is E pv (t), where t represents the hour index, ranging from 1 to 24, and the prediction results of the two models are fused by weighted average: E pv (t)=βE pvtheory (I next (t))+(1-β)Q pred (t) Among them, β is the weight coefficient, which ranges from 0 to 1. It is set to 0.5 in the initial stage and adjusted according to the prediction results. next (t) is the predicted solar radiation intensity at hour t of the next day, Q pred (t) is the power generation at hour t of the next day predicted by the CNN network.

6. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1 is characterized in that: Step S4 determines the charging decision limit of the energy storage system based on the sum of the historical wind power generation and photovoltaic panel power generation and power consumption data, as well as the prediction of future power consumption: Among them, C storage is the rated capacity of the energy storage system, L(t) is the predicted power consumption at hour t, E pv (t is the predicted value of photovoltaic panel power generation, P w (t) is the predicted value of wind turbine power generation.

7. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1 is characterized in that: Step S5 determines the charging decision function. Specifically, when the following conditions are met at the same time, the energy storage system performs charging operation, and the sum of wind power generation and photovoltaic panel power generation is greater than power consumption, that is, P w (t)+E pv (t)-L(t)>0, the energy storage system is not fully charged, that is, S(t) max In this case, the charging power P charge (t) is calculated as follows:​ min means taking the minimum value of the two values ​​in brackets, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at the tth hour, L(t) is the predicted power consumption at the tth hour, which is determined based on the historical power consumption data; η charge is the charging efficiency of the energy storage device, S max is the maximum power allowed by the energy storage system, and S(t) is the power state of the energy storage system at the tth hour; When the market electricity price is low and the above charging conditions are met, in order to obtain more benefits when the electricity price is high, the charging power is appropriately increased. The specific settings are as follows: price (t) <P threshold And P w (t)+E pv (t)-L(t)>0 and S(t) max hour:​ Among them, P price (t) is the known market electricity price at hour t, P threshold It is the set electricity price threshold. When the market electricity price is lower than this threshold, the electricity price is considered to be low, and the condition of increasing the charging power to obtain more profits when the electricity price is high is met. k is the adjustment coefficient, k>1, which is determined by specific market conditions and experience.

8. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1 is characterized in that: Step S6 determines the discharge decision function as follows: when the following conditions are met at the same time, the energy storage system performs a discharge operation, and the sum of the wind power generation and the photovoltaic panel power generation is less than the power consumption, that is, P w (t)+E pv (t)-L(t)<0, there is power available in the energy storage system, that is, S(t)>S min , discharge power P discharge (t) is calculated as follows: min means taking the minimum value, L(t) is the predicted power consumption in the tth hour, P w (t) is the predicted power generation of the wind turbine at hour t, E pv (t) is the predicted power generation of the photovoltaic panel at hour t; η discharge is the discharge efficiency of the energy storage device, S(t) is the state of charge of the energy storage system at the tth hour, S min It is the minimum amount of electricity allowed by the set energy storage system; When the market electricity price is high and the above discharge conditions are met, in order to obtain more benefits, the discharge power is appropriately increased. The specific settings are as follows: price (t)>P threshold2 And P w (t)+E pv When (t)-L(t)<0: Among them, P price (t) is the known market electricity price at hour t, P threshold2 It is the set electricity price threshold. When the market electricity price is higher than this threshold, the electricity price is considered to be high, and the condition for obtaining more profits by increasing the discharge power is met. m is the adjustment coefficient, m>1, which is determined according to market conditions and experience.

9. The economic operation control method for accommodating new energy by distributed energy storage according to claim 8 is characterized in that: When the sum of wind power generation and photovoltaic panel power generation is equal to the power consumption, that is, P w (t)+E pv When (t)-L(t)=0, the energy storage system does not need to be charged or discharged. charge (t) = 0, P discharge (t)=0.

10. The economic operation control method for accommodating new energy by distributed energy storage according to claim 1, characterized in that: Step S7 considers the charging optimization goal of battery life and specifically constructs a mathematical model that reflects the relationship between battery life and charging current, charging depth, and charging temperature factors. The relationship can be expressed as: Where L is the battery life, I c is the charging current, D c is the depth of charge, T c is the charging temperature, L0 is the initial life of the battery, N is the number of charges, a i is a coefficient that measures the relative importance of each factor on the battery life, b is a coefficient related to the number of charging times, reflecting the degree of influence of the increase in charging times on the battery life, and f i It is a function related to each factor, as follows: For the function f1 related to the charging current: Among them, I c is the charging current, I c0 is the reference charging current, specifically the optimal charging current under normal conditions, I cl is the upper limit of the charging current. This function maps the value of the charging current to a specific interval in order to describe its impact on the battery life; For the function f2 related to the depth of charge: Among them, D c is the depth of charge, D c0 It is the reference charging depth, specifically the ideal charging depth when the battery is in normal use, D c1 is the upper limit of the depth of charge. This function also maps the value of the depth of charge to a specific interval to reflect its impact on battery life; For the function f3 related to charging temperature: Among them, T c is the charging temperature, T c0 is the suitable operating temperature of the battery, T c1 is the upper limit of the charging temperature. This function maps the value of the charging temperature to a specific interval to describe its impact on battery life; Use the gradient descent optimization algorithm to gradually find the maximum value of the battery life function by updating the value of the variable along the gradient direction of the function, that is, the direction in which the function changes fastest; Randomly change the charging current from 0 to the maximum allowed I max Initialization charging current I c0 , from 0 to the maximum allowed charging depth D max Initialize the charge depth D c0 And the lowest temperature T at which the battery can work normally min To the maximum temperature T max Initialize charging temperature T c0 , set the learning rate α, the initial value is 0.01, adjust it according to the specific situation, and calculate the partial derivatives of the objective function with respect to each variable for each iteration k; According to the formula of the gradient descent method, update the value of the charging current: in To update the charging current, is the initial charging current, α is the learning rate, is L with respect to the charging current I c The partial derivative of Likewise, update the value of the depth of charge: in To update the charge depth, is the initial charging depth, α is the learning rate, L is about the charge depth D c The partial derivative of Update the value of charging temperature: in Update charging temperature, is the initial charging temperature, α is the learning rate, About charging current I c The partial derivative of When the difference between the updated value of each variable and the value of the previous iteration is less than the set threshold in several consecutive iterations, the algorithm stops iterating. At this time, the current combination of charging current, charging depth and charging temperature is the approximate optimal solution for the longest battery life.

11. An economical operation control system for distributing energy storage to absorb new energy, characterized in that: The method comprises a server and a processor, wherein the server stores a system program, and is characterized in that the processor implements the steps of the method according to any one of claims 1 to 9 when executing the computer program.

Images