A distributed energy storage facility capacity sharing network construction method and device

Abstract

This invention discloses a method and apparatus for constructing a capacity-sharing network for distributed energy storage facilities. The method includes: collecting operational indicators and external environmental data of each distributed energy storage facility and storing them in their respective computers; establishing and training corresponding random forest models in each computer; a central server aggregating the random forest models to obtain a global model and training the global model; each distributed energy storage facility updating its own random forest model according to the global model and sending its current rate of return to the central server; the central server ranking the current rate of return of the distributed energy storage facilities, and sharing the capacity of distributed energy storage facilities with a difference of less than a preset value to form a network, jointly undertaking the energy storage tasks allocated by the power grid. The advantages of this invention are: improving the efficiency of the power grid in allocating energy storage tasks and promoting the efficient operation of distributed energy storage facilities.

Description

Technical Field

[0001] This invention relates to the field of capacity-sharing network construction, and specifically to a method and apparatus for constructing a capacity-sharing network for distributed energy storage facilities. Background Technology

[0002] Factors affecting the profitability of distributed energy storage facilities include weather, operation, operating conditions, events, and environmental factors. However, there is no unified standard for the profitability of distributed energy storage facilities. Due to the lack of cross-sectional comparisons, each distributed energy storage facility finds it difficult to determine its own profitability within its region. Furthermore, submitting relevant data to a central server raises concerns about the leakage of operational data, potentially leading to economic risks.

[0003] Chinese patent publication CN115309647A discloses a privacy-preserving method for software defect prediction based on federated learning. It uses a random forest model built from decision trees for software defect prediction, solving the privacy leakage problem associated with relying on centralized storage of enterprise private data for model training. While it addresses the privacy leakage issue, its primary function is software defect prediction, and it cannot be applied to the power grid to improve the grid's efficiency in allocating energy storage tasks and promote the efficient operation of distributed energy storage facilities. Summary of the Invention

[0004] The technical problem to be solved by this invention is that existing technologies cannot improve the efficiency of grid allocation of energy storage tasks and promote the efficient operation of distributed energy storage facilities.

[0005] This invention solves the above-mentioned technical problems through the following technical means: a method for constructing a distributed energy storage facility capacity sharing network, comprising the following steps:

[0006] S1. Collect the operating indicators and external environmental data of each distributed energy storage facility and store them in their respective computers;

[0007] S2. Build and train the corresponding random forest model on your respective computers;

[0008] S3. The central server aggregates the various random forest models to obtain a global model and trains the global model.

[0009] S4. Each distributed energy storage facility updates its own random forest model according to the global model and sends the current yield to the central server.

[0010] S5. The central server ranks the distributed energy storage facilities according to their current rate of return, and shares the capacity of distributed energy storage facilities with a difference of less than a preset value to form a network, so as to jointly undertake the energy storage tasks assigned by the power grid.

[0011] Further, S1 includes:

[0012] S1.1 In the area requiring diagnosis, all distributed energy storage facilities shall be equipped with distributed acquisition devices, computers with data security functions, and data transmission networks;

[0013] S1.2 Each distributed acquisition device acquires meteorological data, operational data, event data, operating condition data, and atmospheric environment data of the corresponding distributed energy storage facility in real time and stores them in the corresponding computer.

[0014] Further, S2 includes:

[0015] In the computers of each distributed energy storage facility, after reading the real-time data stored locally, the collected parameters are determined as input indicators, and the revenue per unit of energy storage capacity is determined as output indicator. After setting the training samples and validation samples, a random forest model is constructed.

[0016] Furthermore, S2 also includes:

[0017] S2.1 Determine the input and output indicators for each distributed energy storage facility;

[0018] Input indicators include average daily solar irradiance S, average daily temperature T, average daily humidity H, average daily wind speed W, energy storage facility capacity, charging speed, discharging speed, average daily charging times, average daily discharging times, maximum daily charging ratio, and maximum monthly charging ratio; the output indicator is the current rate of return P of the distributed energy storage facility.

[0019] S2.2, Construct a random forest model;

[0020] Distributed energy storage power stations use the local original training dataset S = {x} i ,y i k training sample sets are generated from x in the subset (i = 1, 2, ..., n). i Let y be the input index vector for the i-th sample group. i Let h be the output index vector of the i-th sample group; each training sample set grows independently into a classification tree h without pruning, based on the corresponding training data. i At each node of the tree, randomly select m features from M features (m≤M), and find the optimal feature x from the m features at each node. i(j) Branching growth is performed by selecting a split point for the variable to divide the sample set into two subsets, R1 and R2, where r1 and r2 are the mean responses of the data samples in subsets R1 and R2, respectively. This division is repeated until the total number of data samples in each subset is less than the labeled number. After the division ends, the mean response of the data samples in each subset is output as R1. t For the partitioned subset, r tLet L be the mean response of the data samples in each subset; L is the indicator function, satisfying: The mathematical expression for a regression tree is: Each regression tree model generated by a random forest is independent of the others. The mathematical expression for a random forest model is:

[0021] Further, S3 includes:

[0022] Each distributed energy storage facility uploads its trained random forest model to the central server via a communication network. Without reading the data from each distributed energy storage facility, the central server aggregates the weights using a horizontal federated learning method to complete the global model training for the energy storage facility's profitability.

[0023] Furthermore, the step of using a horizontal federated learning method to aggregate weights and complete the global model training for the profitability of energy storage facilities includes:

[0024] Assume there are N distributed energy storage facilities participating in the model training, and each random forest model uses N datasets for training. ij (i=1,2,3,…,N,j=1,2,3,…,M),F ij Let (w) be the loss function for each set of sample data in the local distributed energy storage facility, then the loss function F for the i-th distributed energy storage facility is... i (w) is:

[0025]

[0026] The total dataset of N distributed energy storage facilities is Loss function of the central server:

[0027]

[0028] The initial parameters for each distributed energy storage facility are w i 0 Each distributed energy storage facility is trained locally, and the local model parameters are updated after each iteration as w. i t The model parameters before iteration are If γ is the global learning rate in federated learning, then...

[0029]

[0030] The central server performs a weighted average of all parameters uploaded by each distributed energy storage facility and updates the global model parameters accordingly.

[0031]

[0032] Where, N iThis represents the weighted value of the parameters uploaded by the i-th distributed energy storage facility. Let represent the gradient of the loss function of the i-th distributed energy storage facility;

[0033] The central server distributes the updated global model parameters to all distributed energy storage facilities as parameter values ​​for the next round of training. When the loss function of the central server reaches a preset value in the global iteration, the iteration stops, and the final global model is obtained.

[0034] Further, S4 includes:

[0035] The central server feeds back the parameter values ​​of the trained global model to each distributed energy storage facility. Each distributed energy storage facility updates its random forest model based on the parameter values ​​of the global model, calculates its current rate of return and next period rate of return, and feeds them back to the central server.

[0036] Further, S5 includes:

[0037] Based on the principle of grouping similar facilities together, the central server sorts distributed energy storage facilities according to the average of the current and next period's rate of return. Distributed energy storage facilities with adjacent average rates of return and a difference of less than a preset value are grouped into 1 tier. Each tier is combined to form a shared energy storage facility, which jointly undertakes the task of energy storage and peak shaving during operation.

[0038] Based on the combination of shared energy storage facilities, a network of interconnected and shared distributed energy storage facilities is drawn. This network is used by power grid companies when allocating energy storage charging and discharging tasks. Power grid companies allocate charging and discharging tasks that match the overall energy storage capacity of the shared energy storage facilities.

[0039] The present invention also provides a device for constructing a distributed energy storage facility capacity sharing network, comprising:

[0040] The data acquisition module is used to collect the operating indicators and external environmental data of each distributed energy storage facility and store them in their respective computers.

[0041] The local model building module is used to build and train the corresponding random forest model on their respective computers.

[0042] The global model building module is used by the central server to aggregate various random forest models, obtain a global model, and train the global model.

[0043] The local model update module is used by each distributed energy storage facility to update its own random forest model according to the global model and send the current yield to the central server.

[0044] The shared network construction module is used by the central server to sort the current rate of return of distributed energy storage facilities, and to share the capacity of distributed energy storage facilities with a difference of less than a preset value to form a network and jointly undertake the energy storage tasks assigned by the power grid.

[0045] Furthermore, the data acquisition module is also used for:

[0046] S1.1 In the area requiring diagnosis, all distributed energy storage facilities shall be equipped with distributed acquisition devices, computers with data security functions, and data transmission networks;

[0047] S1.2 Each distributed acquisition device acquires meteorological data, operational data, event data, operating condition data, and atmospheric environment data of the corresponding distributed energy storage facility in real time and stores them in the corresponding computer.

[0048] Furthermore, the local model building module is also used for:

[0049] In the computers of each distributed energy storage facility, after reading the real-time data stored locally, the collected parameters are determined as input indicators, and the revenue per unit of energy storage capacity is determined as output indicator. After setting the training samples and validation samples, a random forest model is constructed.

[0050] Furthermore, the local model building module is also used for:

[0051] S2.1 Determine the input and output indicators for each distributed energy storage facility;

[0052] Input indicators include average daily solar irradiance S, average daily temperature T, average daily humidity H, average daily wind speed W, energy storage facility capacity, charging speed, discharging speed, average daily charging times, average daily discharging times, maximum daily charging ratio, and maximum monthly charging ratio; the output indicator is the current rate of return P of the distributed energy storage facility.

[0053] S2.2, Construct a random forest model;

[0054] Distributed energy storage power stations use the local original training dataset S = {x} i ,y i k training sample sets are generated from x in the subset (i = 1, 2, ..., n). i Let y be the input index vector for the i-th sample group. i Let h be the output index vector of the i-th sample group; each training sample set grows independently into a classification tree h without pruning, based on the corresponding training data. i At each node of the tree, randomly select m features from M features (m≤M), and find the optimal feature x from the m features at each node. i(j)Branching growth is performed by selecting a split point for the variable to divide the sample set into two subsets, R1 and R2, where r1 and r2 are the mean responses of the data samples in subsets R1 and R2, respectively. This division is repeated until the total number of data samples in each subset is less than the labeled number. After the division ends, the mean response of the data samples in each subset is output as R1. t For the partitioned subset, r t Let L be the mean response of the data samples in each subset; L is the indicator function, satisfying: The mathematical expression for a regression tree is: Each regression tree model generated by a random forest is independent of the others. The mathematical expression for a random forest model is:

[0055] Furthermore, the global model building module is also used for:

[0056] Each distributed energy storage facility uploads its trained random forest model to the central server via a communication network. Without reading the data from each distributed energy storage facility, the central server aggregates the weights using a horizontal federated learning method to complete the global model training for the energy storage facility's profitability.

[0057] Furthermore, the step of using a horizontal federated learning method to aggregate weights and complete the global model training for the profitability of energy storage facilities includes:

[0058] Assume there are N distributed energy storage facilities participating in the model training, and each random forest model uses N datasets for training. ij (i=1,2,3,…,N,j=1,2,3,…,M),F ij Let (w) be the loss function for each set of sample data in the local distributed energy storage facility, then the loss function F for the i-th distributed energy storage facility is... i (w) is:

[0059]

[0060] The total dataset of N distributed energy storage facilities is Loss function of the central server:

[0061]

[0062] The initial parameters for each distributed energy storage facility are W i 0 Each distributed energy storage facility is trained locally, and the local model parameters are updated to W after each iteration. i t The model parameters before iteration are W i t-1 If γ is the global learning rate of federated learning, then we have

[0063] W i t =W i t-1 -γΔF i W i t-1

[0064] The central server performs a weighted average of all parameters uploaded by each distributed energy storage facility and updates the global model parameters accordingly.

[0065]

[0066] Where, N i This represents the weighted value of the parameters uploaded by the i-th distributed energy storage facility. Let represent the gradient of the loss function of the i-th distributed energy storage facility;

[0067] The central server distributes the updated global model parameters to all distributed energy storage facilities as parameter values ​​for the next round of training. When the loss function of the central server reaches a preset value in the global iteration, the iteration stops, and the final global model is obtained.

[0068] Furthermore, the local model update module is also used for:

[0069] The central server feeds back the parameter values ​​of the trained global model to each distributed energy storage facility. Each distributed energy storage facility updates its random forest model based on the parameter values ​​of the global model, calculates its current rate of return and next period rate of return, and feeds them back to the central server.

[0070] Furthermore, the shared network building module is also used for:

[0071] Based on the principle of grouping similar facilities together, the central server sorts distributed energy storage facilities according to the average of the current and next period's rate of return. Distributed energy storage facilities with adjacent average rates of return and a difference of less than a preset value are grouped into 1 tier. Each tier is combined to form a shared energy storage facility, which jointly undertakes the task of energy storage and peak shaving during operation.

[0072] Based on the combination of shared energy storage facilities, a network of interconnected and shared distributed energy storage facilities is drawn. This network is used by power grid companies when allocating energy storage charging and discharging tasks. Power grid companies allocate charging and discharging tasks that match the overall energy storage capacity of the shared energy storage facilities.

[0073] The advantages of this invention are:

[0074] (1) Based on the central server and federated learning technology, this invention utilizes the random forest model transmitted from distributed energy storage facilities to the central server to realize real-time diagnosis of the rate of return of distributed energy storage facilities in the region. The central server sorts the current rate of return of distributed energy storage facilities, shares the capacity of distributed energy storage facilities with similar rankings and forms a network to jointly undertake the energy storage tasks assigned by the power grid, and then assigns energy storage tasks to different energy storage facilities, which helps to improve the efficiency of the power grid in allocating energy storage tasks and promote the efficient operation of distributed energy storage facilities.

[0075] (2) The present invention collects relevant data by the distributed energy storage facility itself and stores it on a secure computer under its control, thereby avoiding the uploading of the power station’s private data and the possible information leakage, effectively ensuring the reliability and confidentiality of the collected data, and providing a reliable data foundation for model training and inefficient diagnosis. Attached Figure Description

[0076] Figure 1 This is a flowchart of a method for constructing a distributed energy storage facility capacity sharing network as disclosed in an embodiment of the present invention;

[0077] Figure 2 This is a schematic diagram of information transmission relationships in a method for constructing a distributed energy storage facility capacity sharing network disclosed in an embodiment of the present invention;

[0078] Figure 3 This is a schematic diagram of the federated learning framework and model training in a method for constructing a distributed energy storage facility capacity sharing network disclosed in an embodiment of the present invention;

[0079] Figure 4 This is a schematic diagram of the capacity sharing network of distributed energy storage facilities in a method for constructing a capacity sharing network of distributed energy storage facilities disclosed in an embodiment of the present invention. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] Example 1

[0082] like Figure 1 and Figure 2 As shown, the present invention provides a method for constructing a distributed energy storage facility capacity sharing network, comprising the following steps:

[0083] S1. Collect operational indicators and external environmental data from each distributed energy storage facility and store them in their respective computers; the specific process is as follows:

[0084] S1.1 In the area requiring diagnosis, all distributed energy storage facilities shall be equipped with distributed acquisition devices, computers with data security functions, and data transmission networks;

[0085] S1.2 Each distributed acquisition device acquires meteorological data, operational data, event data, operating condition data, and atmospheric environment data of the corresponding distributed energy storage facility in real time and stores them in the corresponding secure computer.

[0086] S2. In the computers of each distributed energy storage facility, after reading the real-time data stored locally, determine the collected parameters as input indicators and the revenue per unit of energy storage installed capacity as output indicators. After setting the training samples and validation samples, construct a random forest model. Figure 2 The prediction model in the middle refers to the random forest model corresponding to distributed energy storage facilities. The specific process is as follows:

[0087] S2.1 Determine the input and output indicators for each distributed energy storage facility;

[0088] The input metrics for random forests include, but are not limited to, daily average solar irradiance S, daily average temperature T, daily average humidity H, daily average wind speed W, energy storage facility capacity, charging speed, discharging speed, daily average number of charging cycles, daily average number of discharging cycles, daily maximum charging ratio, and monthly maximum charging ratio; the output metric is the current rate of return P of the distributed energy storage facility.

[0089] S2.2. Determine the collected parameters as input indicators and the current rate of return of energy storage facilities as output indicators, and train a random forest model. Assume there are n sample data points, m input indicators, and x... i Let y be the input index vector for the i-th sample group. i Let be the output index vector of the i-th sample group, and let λ be the weight coefficient vector of the input index. m x0 and y0 are the selected sample data input and output vectors. After setting the training samples and validation samples, the model is trained to obtain a local random forest model with a prediction error rate of less than 5%.

[0090] Random forest is a tree-type classification tree h i The set of (i = 1, 2, ..., n)}. Distributed energy storage power stations use the local original training dataset S = {x i ,y i In the sequence (i = 1, 2, ..., n), bootstrap sampling generates k training sample sets, where x i Let y be the input index vector for the i-th sample group. iLet h be the output index vector of the i-th sample group; each training sample set grows independently into a classification tree h without pruning, based on the corresponding training data. i At each node of the tree, randomly select m features from M features (m≤M), and find the optimal feature x from the m features at each node. i(j) Branching growth is performed by selecting a split point for the variable to divide the sample set into two subsets, R1 and R2, where r1 and r2 are the mean responses of the data samples in subsets R1 and R2, respectively. This division is repeated until the total number of data samples in each subset is less than the labeled number. After the division ends, the mean response of the data samples in each subset is output as R1. t For the partitioned subset, r t This represents the mean response of the data samples in each subset. It should be noted that the regression method for the above classification tree is a mature existing technology and is not an improvement of this invention. For example, the classification tree partitioning method described in the literature "Zhang Songlin. Introduction to CART-Classification and Regression Tree Method [J]. Volcanic Geology and Mineral Resources, 1997(01):67-75." can be used to construct the above classification tree. Therefore, the construction method of the classification tree will not be elaborated upon here.

[0091] L is an indicator function that satisfies:

[0092]

[0093] The mathematical expression for a regression tree is:

[0094]

[0095] Each regression tree model generated by a random forest is independent of the others. The mathematical expression for a random forest model is:

[0096]

[0097] S3, such as Figure 3 As shown, each distributed energy storage facility uploads its trained random forest model to the central server via a communication network. Without reading the data from each distributed energy storage facility, the central server aggregates the weights using a horizontal federated learning method to complete the global model training for the energy storage facility's profitability.

[0098] The method of aggregating weights using a horizontal federated learning approach to train a global model for the profitability of energy storage facilities includes:

[0099] Assume there are N distributed energy storage facilities participating in the model training, and each random forest model uses N datasets for training. ij (i=1,2,3,…,N,j=1,2,3,…,M),F ijLet (w) be the loss function for each set of sample data in the local distributed energy storage facility, then the loss function F for the i-th distributed energy storage facility is... i (w) is:

[0100]

[0101] The total dataset of N distributed energy storage facilities is Loss function of the central server:

[0102]

[0103] The initial parameters for each distributed energy storage facility are W i 0 Each distributed energy storage facility is trained locally, and the local model parameters are updated to W after each iteration. i t The model parameters before iteration are W i t-1 If γ is the global learning rate of federated learning, then we have

[0104]

[0105] The central server performs a weighted average of all parameters uploaded by each distributed energy storage facility and updates the global model parameters accordingly.

[0106]

[0107] Where, N i This represents the weighted value of the parameters uploaded by the i-th distributed energy storage facility. Let represent the gradient of the loss function of the i-th distributed energy storage facility;

[0108] The central server distributes the updated global model parameters to all distributed energy storage facilities as parameter values ​​for the next round of training. When the loss function of the central server reaches a preset value in the global iteration, the iteration stops, and the final global model is obtained.

[0109] S4. The central server feeds back the parameter values ​​of the trained global model to each distributed energy storage facility. Each distributed energy storage facility updates its random forest model based on the parameter values ​​of the global model, calculates its current rate of return and next period rate of return, and feeds them back to the central server.

[0110] S5, such as Figure 4As shown, the central server, based on the principle of grouping similar facilities together, sorts distributed energy storage facilities according to the average of the current and next period's returns. Distributed energy storage facilities with similar average returns are grouped into one tier, and each tier is combined to form a shared energy storage facility, which jointly undertakes energy storage and peak shaving tasks during operation. For example, if the average returns are 0.8, 0.81, 0.81, 0.82, 0.82, 0.85, 0.85, 0.86, and 0.90, the distributed energy storage facilities with average returns of 0.8, 0.81, 0.81, 0.82, and 0.82 are grouped into one tier, the distributed energy storage facilities with average returns of 0.85, 0.85, and 0.86 are grouped into another tier, and the distributed energy storage facility with a return of 0.90 is set up separately and not grouped.

[0111] Based on the combination of shared energy storage facilities, a network of interconnected and shared distributed energy storage facilities is drawn. This network is used by power grid companies when allocating energy storage charging and discharging tasks. Power grid companies allocate charging and discharging tasks that match the overall energy storage capacity of the shared energy storage facilities.

[0112] Through the above technical solutions, this invention, based on a central server and federated learning technology, utilizes a random forest model transmitted from distributed energy storage facilities to the central server to achieve real-time diagnosis of the profitability of distributed energy storage facilities within a region. The central server ranks the current profitability of distributed energy storage facilities, shares the capacity of distributed energy storage facilities with similar rankings, and forms a network to jointly undertake the energy storage tasks assigned by the power grid. This, in turn, allocates energy storage tasks to different energy storage facilities, which helps improve the efficiency of the power grid in allocating energy storage tasks and promotes the efficient operation of distributed energy storage facilities.

[0113] Example 2

[0114] Based on Embodiment 1, Embodiment 2 of the present invention also provides a distributed energy storage facility capacity sharing network construction device, comprising:

[0115] The data acquisition module is used to collect the operating indicators and external environmental data of each distributed energy storage facility and store them in their respective computers.

[0116] The local model building module is used to build and train the corresponding random forest model on their respective computers.

[0117] The global model building module is used by the central server to aggregate various random forest models, obtain a global model, and train the global model.

[0118] The local model update module is used by each distributed energy storage facility to update its own random forest model according to the global model and send the current yield to the central server.

[0119] The shared network construction module is used by the central server to sort the current rate of return of distributed energy storage facilities, and to share the capacity of distributed energy storage facilities with a difference of less than a preset value to form a network and jointly undertake the energy storage tasks assigned by the power grid.

[0120] Specifically, the data acquisition module is also used for:

[0121] S1.1 In the area requiring diagnosis, all distributed energy storage facilities shall be equipped with distributed acquisition devices, computers with data security functions, and data transmission networks;

[0122] S1.2 Each distributed acquisition device acquires meteorological data, operational data, event data, operating condition data, and atmospheric environment data of the corresponding distributed energy storage facility in real time and stores them in the corresponding computer.

[0123] Specifically, the local model building module is also used for:

[0124] In the computers of each distributed energy storage facility, after reading the real-time data stored locally, the collected parameters are determined as input indicators, and the revenue per unit of energy storage capacity is determined as output indicator. After setting the training samples and validation samples, a random forest model is constructed.

[0125] More specifically, the local model building module is also used for:

[0126] S2.1 Determine the input and output indicators for each distributed energy storage facility;

[0127] Input indicators include average daily solar irradiance S, average daily temperature T, average daily humidity H, average daily wind speed W, energy storage facility capacity, charging speed, discharging speed, average daily charging times, average daily discharging times, maximum daily charging ratio, and maximum monthly charging ratio; the output indicator is the current rate of return P of the distributed energy storage facility.

[0128] S2.2, Construct a random forest model;

[0129] Distributed energy storage power stations use the local original training dataset S = {x} i ,y i k training sample sets are generated from x in the subset (i = 1, 2, ..., n). i Let y be the input index vector for the i-th sample group. i Let h be the output index vector of the i-th sample group; each training sample set grows independently into a classification tree h without pruning, based on the corresponding training data. i At each node of the tree, randomly select m features from M features (m≤M), and find the optimal feature x from the m features at each node. i(j)Branching growth is performed by selecting a split point for the variable to divide the sample set into two subsets, R1 and R2, where r1 and r2 are the mean responses of the data samples in subsets R1 and R2, respectively. This division is repeated until the total number of data samples in each subset is less than the labeled number. After the division ends, the mean response of the data samples in each subset is output as R1. t For the partitioned subset, r t Let L be the mean response of the data samples in each subset; L is the indicator function, satisfying: The mathematical expression for a regression tree is: Each regression tree model generated by a random forest is independent of the others. The mathematical expression for a random forest model is:

[0130] Specifically, the global model building module is also used for:

[0131] Each distributed energy storage facility uploads its trained random forest model to the central server via a communication network. Without reading the data from each distributed energy storage facility, the central server aggregates the weights using a horizontal federated learning method to complete the global model training for the energy storage facility's profitability.

[0132] More specifically, the method of aggregating weights using a horizontal federated learning approach to train a global model for the profitability of energy storage facilities includes:

[0133] Assume there are N distributed energy storage facilities participating in the model training, and each random forest model uses N datasets for training. ij (i=1,2,3,…,N,j=1,2,3,…,M),F ij Let (w) be the loss function for each set of sample data in the local distributed energy storage facility, then the loss function F for the i-th distributed energy storage facility is... i (w) is:

[0134]

[0135] The total dataset of N distributed energy storage facilities is Loss function of the central server:

[0136]

[0137] The initial parameters for each distributed energy storage facility are W i 0 Each distributed energy storage facility is trained locally, and the local model parameters are updated to W after each iteration. i t The model parameters before iteration are W i t-1 If γ is the global learning rate of federated learning, then we have

[0138] W i t =W i t-1 -γΔF i W i t-1

[0139] The central server performs a weighted average of all parameters uploaded by each distributed energy storage facility and updates the global model parameters accordingly.

[0140]

[0141] Where, N i This represents the weighted value of the parameters uploaded by the i-th distributed energy storage facility. Let represent the gradient of the loss function of the i-th distributed energy storage facility;

[0142] The central server distributes the updated global model parameters to all distributed energy storage facilities as parameter values ​​for the next round of training. When the loss function of the central server reaches a preset value in the global iteration, the iteration stops, and the final global model is obtained.

[0143] Specifically, the local model update module is also used for:

[0144] The central server feeds back the parameter values ​​of the trained global model to each distributed energy storage facility. Each distributed energy storage facility updates its random forest model based on the parameter values ​​of the global model, calculates its current rate of return and next period rate of return, and feeds them back to the central server.

[0145] Specifically, the shared network construction module is also used for:

[0146] Based on the principle of grouping similar facilities together, the central server sorts distributed energy storage facilities according to the average of the current and next period's rate of return. Distributed energy storage facilities with adjacent average rates of return and a difference of less than a preset value are grouped into 1 tier. Each tier is combined to form a shared energy storage facility, which jointly undertakes the task of energy storage and peak shaving during operation.

[0147] Based on the combination of shared energy storage facilities, a network of interconnected and shared distributed energy storage facilities is drawn. This network is used by power grid companies when allocating energy storage charging and discharging tasks. Power grid companies allocate charging and discharging tasks that match the overall energy storage capacity of the shared energy storage facilities.

[0148] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a distributed energy storage facility capacity-sharing network, characterized in that, Includes the following steps: S1. Collect the operating indicators and external environmental data of each distributed energy storage facility and store them in their respective computers; S2. Build and train the corresponding random forest model on your respective computers; S3. The central server aggregates the various random forest models to obtain a global model and trains the global model. Each distributed energy storage facility uploads its trained random forest model to the central server via a communication network. Without reading the data from each distributed energy storage facility, the central server aggregates the weights using a horizontal federated learning method to complete the global model training for the energy storage facility's profitability. The method of aggregating weights using a horizontal federated learning approach to train a global model for the profitability of energy storage facilities includes: Assume there are N distributed energy storage facilities participating in the model training, and the dataset used for training each random forest model is... , Let be the loss function for each set of sample data at the local location of the distributed energy storage facility. Then, the loss function for the i-th distributed energy storage facility... for: The total dataset of N distributed energy storage facilities is The loss function of the central server: ； The initial parameters for each distributed energy storage facility are as follows: Each distributed energy storage facility is trained locally, and the local model parameters are updated after each iteration. The model parameters before iteration are , If it is the global learning rate of federated learning, then... The central server performs a weighted average of all parameters uploaded by each distributed energy storage facility and updates the global model parameters accordingly. in, This represents the weighted value of the parameters uploaded by the i-th distributed energy storage facility. Let represent the gradient of the loss function of the i-th distributed energy storage facility; The central server distributes the updated global model parameters to all distributed energy storage facilities as parameter values ​​for the next round of training. When the loss function of the central server in the global iteration reaches a preset value, the iteration stops and the final global model is obtained. S4. Each distributed energy storage facility updates its own random forest model according to the global model and sends the current yield to the central server. S5. The central server ranks the distributed energy storage facilities according to their current rate of return, and shares the capacity of distributed energy storage facilities with a difference of less than a preset value to form a network, so as to jointly undertake the energy storage tasks assigned by the power grid.

2. The method for constructing a distributed energy storage facility capacity sharing network according to claim 1, characterized in that, S1 includes: S1.1 In the area requiring diagnosis, all distributed energy storage facilities shall be equipped with distributed acquisition devices, computers with data security functions, and data transmission networks; S1.2 Each distributed acquisition device acquires meteorological data, operational data, event data, operating condition data, and atmospheric environment data of the corresponding distributed energy storage facility in real time and stores them in the corresponding computer.

3. The method for constructing a distributed energy storage facility capacity sharing network according to claim 1, characterized in that, S2 includes: In the computers of each distributed energy storage facility, after reading the real-time data stored locally, the collected parameters are determined as input indicators, and the revenue per unit of energy storage capacity is determined as output indicator. After setting the training samples and validation samples, a random forest model is constructed.

4. The method for constructing a distributed energy storage facility capacity sharing network according to claim 3, characterized in that, S2 further includes: S2.1 Determine the input and output indicators for each distributed energy storage facility; Input indicators include average daily solar irradiance S, average daily temperature T, average daily humidity H, average daily wind speed W, energy storage facility capacity, charging speed, discharging speed, average daily charging times, average daily discharging times, maximum daily charging ratio, and maximum monthly charging ratio; the output indicator is the current rate of return P of the distributed energy storage facility. S2.2, Construct a random forest model; Distributed energy storage power stations use local raw training datasets k training sample sets are generated by sampling from the middle, where, For the first i The input index vector of the group sample. For the first i The output index vector of the group of samples; each training sample set grows independently into a regression tree without pruning based on the corresponding training data. At each node of the tree, randomly select m features from M features (m≤M), and find the optimal feature from the m features at each node. Branching growth is performed, i.e., selecting the optimal feature. The split point divides the sample set into two subsets. and , and For subset and The mean response of the data samples in the dataset is used as the output. This process is repeated until the total number of data samples in each subset is less than the specified number. After the partitioning is complete, the mean response of the data samples in each subset is used as the output. For the partitioned subset, This represents the mean response of the data samples in each subset; Let be an indicator function, satisfying: The mathematical expression for a regression tree is: Each regression tree model generated by a random forest is independent of the others. The mathematical expression for a random forest model is: .

5. The method for constructing a distributed energy storage facility capacity sharing network according to claim 1, characterized in that, S4 includes: The central server feeds back the parameter values ​​of the trained global model to each distributed energy storage facility. Each distributed energy storage facility updates its random forest model based on the parameter values ​​of the global model, calculates its current rate of return and next period rate of return, and feeds them back to the central server.

6. The method for constructing a distributed energy storage facility capacity sharing network according to claim 1, characterized in that, S5 includes: Based on the principle of grouping similar facilities together, the central server sorts distributed energy storage facilities according to the average of the current and next period's rate of return. Distributed energy storage facilities with adjacent average rates of return and a difference of less than a preset value are grouped into 1 tier. Each tier is combined to form a shared energy storage facility, which jointly undertakes the task of energy storage and peak shaving during operation. Based on the combination of shared energy storage facilities, a network of interconnected and shared distributed energy storage facilities is drawn. This network is used by power grid companies when allocating energy storage charging and discharging tasks. Power grid companies allocate charging and discharging tasks that match the overall energy storage capacity of the shared energy storage facilities.

7. A device for constructing a distributed energy storage facility capacity sharing network, characterized in that, include: The data acquisition module is used to collect the operating indicators and external environmental data of each distributed energy storage facility and store them in their respective computers. The local model building module is used to build and train the corresponding random forest model on their respective computers. The global model building module is used by the central server to aggregate various random forest models, obtain a global model, and train the global model. Each distributed energy storage facility uploads its trained random forest model to the central server via a communication network. Without reading the data from each distributed energy storage facility, the central server aggregates the weights using a horizontal federated learning method to complete the global model training for the energy storage facility's profitability. The method of aggregating weights using a horizontal federated learning approach to train a global model for the profitability of energy storage facilities includes: Assume there are N distributed energy storage facilities participating in the model training, and the dataset used for training each random forest model is... , Let be the loss function for each set of sample data at the local location of the distributed energy storage facility. Then, the loss function for the i-th distributed energy storage facility... for: The total dataset of N distributed energy storage facilities is The loss function of the central server: ； The initial parameters for each distributed energy storage facility are as follows: Each distributed energy storage facility is trained locally, and the local model parameters are updated after each iteration. The model parameters before iteration are , If it is the global learning rate of federated learning, then... The central server performs a weighted average of all parameters uploaded by each distributed energy storage facility and updates the global model parameters accordingly. in, This represents the weighted value of the parameters uploaded by the i-th distributed energy storage facility. Let represent the gradient of the loss function of the i-th distributed energy storage facility; The central server distributes the updated global model parameters to all distributed energy storage facilities as parameter values ​​for the next round of training. When the loss function of the central server in the global iteration reaches a preset value, the iteration stops and the final global model is obtained. The local model update module is used by each distributed energy storage facility to update its own random forest model according to the global model and send the current yield to the central server. The shared network construction module is used by the central server to sort the current rate of return of distributed energy storage facilities, and to share the capacity of distributed energy storage facilities with a difference of less than a preset value to form a network and jointly undertake the energy storage tasks assigned by the power grid.

8. The distributed energy storage facility capacity sharing network construction device according to claim 7, characterized in that, The data acquisition module is also used for: S1.1 In the area requiring diagnosis, all distributed energy storage facilities shall be equipped with distributed acquisition devices, computers with data security functions, and data transmission networks; S1.2 Each distributed acquisition device acquires meteorological data, operational data, event data, operating condition data, and atmospheric environment data of the corresponding distributed energy storage facility in real time and stores them in the corresponding computer.

Images