Adjustable potential prediction method suitable for power consumption side distributed resources and related equipment

Abstract

The application discloses a method for predicting adjustable potential of distributed resources on a power consumption side and related equipment. The method comprises the following steps: obtaining historical adjustable potential data of distributed resources; obtaining a pre-constructed adjustable potential prediction cascaded time domain memory network model; training the adjustable potential prediction cascaded time domain memory network model based on the historical adjustable potential data by using a residual learning algorithm to obtain an optimal adjustable potential prediction cascaded time domain memory network model, wherein the adjustable potential prediction cascaded time domain memory network model adopts a multi-sub-memory storage area cascaded layer structure to realize cross-layer feature fusion; and inputting real-time adjustable potential data of the distributed resources into the optimal adjustable potential prediction cascaded time domain memory network model to obtain a predicted value of the adjustable potential data. The application can realize deep fusion of multi-source data, improve prediction accuracy and prediction stability, and meet the real-time and reliability requirements of power grid dispatching for adjustable potential prediction.

Description

Technical Field

[0001] This invention belongs to the field of power system dispatching and distributed resource management technology, specifically relating to a method and related equipment for predicting the adjustable potential of distributed resources on the power consumption side. Background Technology

[0002] With the rapid development of new energy technologies, the penetration rate of distributed resources on the electricity consumption side continues to increase, and their adjustable potential has become a key resource for peak shaving and valley filling in the power grid to ensure power supply stability. Accurately predicting the adjustable potential of distributed resources on the electricity consumption side is a core prerequisite for achieving optimized scheduling of distributed resources and improving the operating efficiency of the power grid.

[0003] In existing technologies, methods for predicting the adjustable potential of distributed resources on the power consumption side mainly include traditional statistical methods and conventional machine learning methods. However, these methods have significant drawbacks: traditional statistical methods struggle to handle multi-dimensional, strongly time-correlated sampled data and cannot effectively integrate multi-source features such as adjustable power, load values, and demand response signals, thus limiting prediction accuracy; conventional machine learning methods often employ single-layer network structures, lacking cross-layer feature interaction mechanisms and failing to capture complex time-series patterns; insufficient weight optimization during model training easily leads to overfitting or slow convergence, and the lack of effective stability determination mechanisms results in large fluctuations in prediction results.

[0004] Therefore, there is an urgent need for an adjustable potential prediction method and related equipment that can meet the requirements of multi-dimensional time series characteristics and ensure prediction stability and accuracy. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method and related equipment for predicting the adjustable potential of distributed resources on the power consumption side. This method enables deep fusion of multi-source data, improves prediction accuracy and model stability, and meets the real-time and reliability requirements of power grid dispatch for adjustable potential prediction.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting the adjustable potential of distributed resources on the electricity consumption side, comprising:

[0008] Obtain historical adjustable potential data of distributed resources;

[0009] A pre-constructed adjustable potential prediction cascaded temporal memory network model is obtained. Based on historical adjustable potential data, the adjustable potential prediction cascaded temporal memory network model is trained using the residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model. The adjustable potential prediction cascaded temporal memory network model adopts a multi-sub-memory storage area cascaded stacked structure to achieve cross-level feature fusion.

[0010] The real-time adjustable potential data of the acquired distributed resources is input into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data.

[0011] In conjunction with the first aspect, optionally, the adjustable potential sampling data of the distributed resources includes adjustable power, load values ​​on the electricity consumption side, and demand responses issued by the grid side;

[0012] The adjustable power is represented by a matrix. express:

[0013] ,

[0014] In the formula, This represents the actual adjustable power during time period t;

[0015] The power-side load value is represented by a matrix. express:

[0016] ,

[0017] In the formula, This represents the total load on the electricity consumption side during time period t;

[0018] The demand response issued by the power grid side express:

[0019] ,

[0020] In the formula, This represents the demand response value issued by the power grid side during time period t. This represents the device number of the demand response equipment issued by the power grid side during time period t.

[0021] In conjunction with the first aspect, optionally, the adjustable potential prediction cascaded temporal memory network model is constructed through the following steps:

[0022] Step 1: Initialize the network parameters of the basic temporal memory network, which includes a feature input layer, a memory storage layer, and a temporal output layer arranged sequentially; the network parameters include the input weight matrix of the feature input layer. The internal weight matrix of the memory storage layer Output weight matrix of the time-domain output layer ,in, This indicates the input dimension of the feature input layer. This indicates the number of neurons in the memory storage layer. Indicates the output dimension of the time-domain output layer;

[0023] Step 2: Cascade the memory storage layers of multiple basic temporal memory networks into cascaded sub-memory storage areas. Each memory storage layer serves as a sub-memory storage area within the cascaded sub-memory storage area. Various types of adjustable potential sampling data are mapped to the input vectors of the corresponding sub-memory storage areas, and the dimension of the input vectors matches the dimension of the input weight matrix of the corresponding sub-memory storage area.

[0024] The cascaded sub-memory storage area comprises K sub-memory storage areas, and the cross-level coupling weight matrix between the k-th sub-memory storage area and the m-th sub-memory storage area is:

[0025] ,

[0026] The real-time state of the m-th sub-memory storage area is transferred to the k-th sub-memory storage area to achieve cross-level feature fusion. The state vector update equation of the k-th sub-memory storage area is:

[0027] ,

[0028] In the formula, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents a constant between 1 and 0, called the leakage rate, which controls how quickly the current state forgets past states; This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents the input weight matrix of the k-th basic temporal memory network feature input layer. This indicates that the k-th basic temporal memory network feature input layer is in the k-th... The input vector for the time period, Its function is to input the k-th basic temporal memory network feature into the layer at the th ... Input vector for a given time period Mapped to the neuron space of the k-th sub-memory storage area; This represents the internal weight matrix of the k-th sub-memory storage area, i.e., the connection weights between neurons within the same sub-memory storage area, used to transmit the historical state of that sub-memory storage area; This indicates that the m-th sub-memory storage area is in the... State vector for a given time period; coupling terms The m-th sub-memory storage area refers to the m-th sub-memory storage area in the first... The state vector of a time period is coupled with a weight matrix across levels. The feature interaction term, after mapping, is passed to the k-th sub-memory storage area to achieve deep fusion of data features at different levels; tanh is the activation function.

[0029] Step 3: Concatenate the time-domain output layers of multiple basic time-domain memory networks to obtain an adjustable potential prediction cascaded time-domain memory network model.

[0030] In conjunction with the first aspect, optionally, the state update formula for each basic temporal memory network is:

[0031] ,

[0032] The formula for calculating its output data is:

[0033] ,

[0034] in, This represents the state vector of the memory storage layer at time t+1. This represents the state vector of the memory storage layer at time t. This represents the input vector of the feature input layer of the basic temporal memory network at time t. This represents the output vector of the temporal output layer of the basic temporal memory network at time t. This represents the bias vector. and Both represent activation functions.

[0035] In conjunction with the first aspect, optionally, the step of training the adjustable potential prediction cascaded temporal memory network model using the residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model includes:

[0036] Step 1: Initialize the relevant parameters of the residual learning algorithm and the weight matrices of each layer of the adjustable potential prediction cascaded temporal memory network model. The number of sub-memory storage areas is... Each sub-memory storage area contains Q neurons, and the maximum number of iterations is T;

[0037] Step 2: Input the prepared historical adjustable potential data into the constructed cascaded temporal memory network model for forward propagation. For the k-th basic temporal memory network, update the memory storage area state according to the following formula:

[0038] ,

[0039] in, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period This indicates that the k-th basic temporal memory network feature input layer is in the k-th... The input vector for the time period; This represents the state vector of the k-th sub-memory storage area in time period t;

[0040] Step 3: Based on the state vector of the k-th sub-memory storage area Calculate the output of the temporal output layer of the k-th basic temporal memory network:

[0041] ,

[0042] in, This represents the output of the temporal output layer of the k-th basic temporal memory network. This represents the output weight matrix of the k-th basic temporal memory network;

[0043] Ultimately, the basic output of the entire adjustable potential prediction cascaded temporal memory network model is... This is the set of outputs from the temporal output layers of each basic temporal memory network:

[0044] ,

[0045] in, This represents the output of the temporal output layer of the i-th adjustable potential sampled data at the k-th basic temporal memory network;

[0046] Step 4: Use the mean of squared deviations as the loss function to measure the difference between the network's predicted output and the actual expected output. The loss function is:

[0047] ,

[0048] Where N represents the number of adjustable potential samples. This represents the actual output of the i-th adjustable potential sampled data. This represents the predicted output of the i-th adjustable potential sampled data. This represents the l2 norm, used to calculate the square root of the sum of squares of vector elements;

[0049] Step 5: To compensate for the prediction bias of the base output, a margin function is introduced to construct a margin connection mechanism, the final prediction output of which is:

[0050] ,

[0051] in, For all sub-memory storage areas in the first The concatenated state vector is formed by concatenating the updated state vectors for each time period. This is a nonlinear residual function used to extract bias compensation information from the global state of the sub-memory storage area, and is ultimately used as the base output of the cascaded temporal memory network model with adjustable potential prediction. The superposition of these values ​​constitutes the predicted values ​​of the optimal adjustable potential prediction cascade temporal memory network model. ;

[0052] Step 6: Perform backpropagation based on the residual learning algorithm to solve for the gradient of the loss function with respect to each weight matrix; whereby, when updating the output weight matrix from the memory storage area to the time-domain output layer... At that time, according to the chain rule, the loss function right The gradient is:

[0053] ,

[0054] in, Indicates that the sub-memory storage area is in the first... The updated state vector for each time period; the loss function applied to the input weight matrix. and internal weight matrix To obtain the gradient, it is necessary to traverse all sub-memory storage areas and accumulate the gradient contribution of each storage area:

[0055] ,

[0056] ,

[0057] Step 7: Update the input weight matrix using the residual gradient iteration method. Internal weight matrix and output weight matrix ;

[0058] Determine if the current loss function has converged; if not, proceed to step 2; if yes, proceed to step 8.

[0059] Step 8: Output the global optimal parameter set The training of the prediction network is completed, resulting in the optimal adjustable potential prediction cascade temporal memory network model, and the globally optimal parameter set is obtained. Including the optimal input weight matrix Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector .

[0060] In conjunction with the first aspect, optionally, the residual gradient iteration method can be used to update the input weight matrix. Internal weight matrix and output weight matrix The update formula used is:

[0061] ,

[0062] ,

[0063] ,

[0064] in, , , These are the updated input weight matrix, internal weight matrix, and output weight matrix, respectively. It is the learning rate, used to control the step size of each parameter update.

[0065] In conjunction with the first aspect, the adjustable potential prediction method may optionally further include:

[0066] The stability of the adjustable potential prediction cascade temporal memory network model optimized by residual learning is determined using the following formula:

[0067] ,

[0068] in, This represents the internal weight matrix of the k-th sub-memory storage area optimized through margin learning. Indicates the spectral radius.

[0069] In conjunction with the first aspect, optionally, the acquired real-time adjustable potential data of the distributed resources is input into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data:

[0070] Step 1: Obtain data related to the adjustable potential of distributed resources within the new sampling period, including the newly sampled adjustable power matrix. Electricity load matrix and the demand response signal matrix issued by the power grid side :

[0071] ,

[0072] in, For newly sampled adjustable potential data;

[0073] Step 2: Call the globally optimal parameter set obtained during the training phase, including the optimal input weight matrix. Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector The network parameters for the optimal adjustable potential prediction cascaded temporal memory network are set, and the initial state of each sub-memory storage area is set as the final state vector after convergence during the training phase. ;

[0074] Step 3: Newly sampled adjustable potential data The data is sequentially input into the optimal adjustable potential prediction cascaded temporal memory network according to the time series, and forward propagation calculation is performed. The state of each sub-memory storage area is updated in real time. For the k-th sub-memory storage area, the state update equation for time period t is:

[0075] ,

[0076] in, Indicates the first The optimal state vector of each sub-memory storage area in time period t. Indicates the first The sub-memory storage area is in the first The state vector of a time period This represents the input vector of the new adjustable potential sampling data in time period t. Indicates the first Each sub-memory storage area for the first Cross-level feature coupling terms in individual memory storage areas;

[0077] Step 4: Concatenate the optimal state vectors of all sub-memory storage areas at time t to obtain the concatenated state vector. :

[0078] ,

[0079] In the formula, For the first The optimal state vector of each sub-memory storage area in time period t;

[0080] Based on the optimization results of the residual learning algorithm, the predicted value for time period t is calculated through the time-domain output layer:

[0081] ,

[0082] in, This represents the predicted value of the adjustable potential data of distributed resources on the electricity consumption side during time period t. Represents the state concatenation vector The diagonalized matrix;

[0083] Step 5: Repeat steps 3 and 4 to complete the adjustable potential prediction for all time periods within the new sampling period, and output the final adjustable potential prediction result matrix:

[0084] .

[0085] Secondly, the present invention provides an adjustable potential prediction device suitable for distributed resources on the electricity consumption side, comprising:

[0086] The historical adjustable potential data acquisition module is used to acquire historical adjustable potential data of distributed resources;

[0087] The module for obtaining the optimal adjustable potential prediction cascaded temporal memory network model is used to acquire a pre-constructed adjustable potential prediction cascaded temporal memory network model. Based on historical adjustable potential data, the module trains the adjustable potential prediction cascaded temporal memory network model using a residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model. The adjustable potential prediction cascaded temporal memory network model adopts a multi-sub-memory storage area cascaded stacked structure to achieve cross-level feature fusion.

[0088] The prediction module is used to input the real-time adjustable potential data of the acquired distributed resources into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data.

[0089] Thirdly, the present invention provides an adjustable potential prediction system suitable for distributed resources on the power consumption side, including a storage medium and a processor;

[0090] The storage medium is used to store instructions;

[0091] The processor is configured to operate according to the instructions to perform the method according to any one of claims 1-8.

[0092] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0093] This invention achieves deep interaction of multi-source data such as adjustable power, load values, and demand response signals through the multi-sub-region structure of a cascaded time-domain memory network and cross-level feature fusion. Combined with a margin learning algorithm to optimize the weight matrix, it significantly reduces prediction errors. A spectral radius determination mechanism is introduced to ensure stable network operation, and a leakage rate parameter controls the state forgetting speed, improving the model's adaptability to data volatility during model training. The model exhibits fast convergence speed and high forward propagation computation efficiency, meeting the real-time prediction requirements of the adjustable potential of distributed resources on the power consumption side. It is applicable to various types of distributed resources on the power consumption side and can flexibly adapt to different sampling periods and data dimensions. Attached Figure Description

[0094] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0095] Figure 1 This is a schematic diagram of a process for constructing an adjustable potential prediction cascaded temporal memory network model provided by an embodiment of the present invention;

[0096] Figure 2 This is a training flowchart provided by an embodiment of the present invention for obtaining the optimal parameter set by training a network model using a residual learning algorithm;

[0097] Figure 3 This is a schematic diagram of a process for predicting the adjustable potential of distributed resources on the power consumption side using an optimal adjustable potential prediction cascaded time-domain memory network model, as provided in an embodiment of the present invention. Detailed Implementation

[0098] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0099] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0100] Example 1

[0101] This invention provides a method for predicting the adjustable potential of distributed resources on the power consumption side, comprising the following steps:

[0102] (1) Obtain historical adjustable potential data of distributed resources;

[0103] (2) Obtain a pre-constructed adjustable potential prediction cascaded temporal memory network model. Based on historical adjustable potential data, train the adjustable potential prediction cascaded temporal memory network model using the residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model. The adjustable potential prediction cascaded temporal memory network model adopts a multi-sub-memory storage area cascaded stacked structure to achieve cross-level feature fusion.

[0104] (3) Input the real-time adjustable potential data of the acquired distributed resources into the optimal adjustable potential prediction cascade time-domain memory network model to obtain the predicted value of the adjustable potential data.

[0105] In one specific embodiment of the present invention, the adjustable potential sampling data of the distributed resources includes adjustable power, load value on the electricity consumption side, and demand response issued by the grid side;

[0106] The adjustable power is represented by a matrix. express:

[0107] ,

[0108] In the formula, This represents the actual adjustable power during time period t;

[0109] The power-side load value is represented by a matrix. express:

[0110] ,

[0111] In the formula, This represents the total load on the electricity consumption side during time period t;

[0112] The demand response issued by the power grid side express:

[0113] ,

[0114] In the formula, This represents the demand response value issued by the power grid side during time period t. This represents the device number of the demand response equipment issued by the power grid side during time period t.

[0115] In one specific embodiment of the present invention, the adjustable potential prediction cascaded temporal memory network model is constructed through the following steps:

[0116] Step 1: Initialize the network parameters of the basic temporal memory network, which includes a feature input layer, a memory storage layer, and a temporal output layer arranged sequentially; the network parameters include the input weight matrix of the feature input layer. The internal weight matrix of the memory storage layer Output weight matrix of the time-domain output layer ,in, This indicates the input dimension of the feature input layer. This indicates the number of neurons in the memory storage layer. Indicates the output dimension of the time-domain output layer;

[0117] Step 2: Cascade the memory storage layers of multiple basic temporal memory networks into cascaded sub-memory storage areas. Each memory storage layer serves as a sub-memory storage area within the cascaded sub-memory storage area. Various types of adjustable potential sampling data are mapped to the input vectors of the corresponding sub-memory storage areas, and the dimension of the input vectors matches the dimension of the input weight matrix of the corresponding sub-memory storage area.

[0118] The cascaded sub-memory storage area comprises K sub-memory storage areas, and the cross-level coupling weight matrix between the k-th sub-memory storage area and the m-th sub-memory storage area is:

[0119] ,

[0120] The real-time state of the m-th sub-memory storage area is transferred to the k-th sub-memory storage area to achieve cross-level feature fusion. The state vector update equation of the k-th sub-memory storage area is:

[0121] ,

[0122] In the formula, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents a constant between 1 and 0, called the leakage rate, which controls how quickly the current state forgets past states; This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents the input weight matrix of the k-th basic temporal memory network feature input layer. This represents the input vector of the k-th basic temporal memory network feature input layer. Its function is to input the feature vector of the k-th basic temporal memory network into the input layer. Mapped to the neuron space of the k-th sub-memory storage area; This represents the internal weight matrix of the k-th sub-memory storage area, i.e., the connection weights between neurons within the same sub-memory storage area, used to transmit the historical state of that sub-memory storage area; This indicates that the m-th sub-memory storage area is in the... State vector for a given time period; coupling terms The m-th sub-memory storage area refers to the m-th sub-memory storage area in the... The state vector of a time period is coupled with a weight matrix across levels. After mapping, the feature interaction term is passed to the k-th sub-memory storage area to achieve deep fusion of data features at different levels; tanh is the activation function.

[0123] Step 3: Concatenate the time-domain output layers of multiple basic time-domain memory networks to obtain an adjustable potential prediction cascaded time-domain memory network model.

[0124] In one specific embodiment of the present invention, the state update formula for each basic temporal memory network is as follows:

[0125] ,

[0126] The formula for calculating its output data is:

[0127] ,

[0128] in, This represents the state vector of the memory storage layer at time t+1. This represents the state vector of the memory storage layer at time t. This represents the input vector of the feature input layer at time t. This represents the output vector of the temporal output layer of the basic temporal memory network at time t. This represents the bias vector. and Both represent activation functions.

[0129] In one specific embodiment of the present invention, training the adjustable potential prediction cascaded temporal memory network model using a residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model includes:

[0130] Step 1: Initialize the relevant parameters of the residual learning algorithm and the weight matrices of each layer of the adjustable potential prediction cascaded temporal memory network model. The number of sub-memory storage areas is... Each sub-memory storage area contains Q neurons, and the maximum number of iterations is T;

[0131] Step 2: Input the prepared historical adjustable potential data into the constructed cascaded temporal memory network model for forward propagation. For the k-th basic temporal memory network, update the memory storage area state according to the following formula:

[0132] ,

[0133] in, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period This represents the input of the k-th sub-memory storage area in time period t; This represents the state vector of the k-th sub-memory storage area in time period t;

[0134] Step 3: Based on the state vector of the k-th sub-memory storage area Calculate the output of the temporal output layer of the k-th basic temporal memory network:

[0135] ,

[0136] in, This represents the output of the temporal output layer of the k-th basic temporal memory network. This represents the output weight matrix of the k-th basic temporal memory network;

[0137] Ultimately, the basic output of the entire adjustable potential prediction cascaded temporal memory network model is... This is the set of outputs from the temporal output layers of each basic temporal memory network:

[0138] ,

[0139] in, This represents the output of the temporal output layer of the i-th adjustable potential sampled data at the k-th basic temporal memory network;

[0140] Step 4: Use the mean of squared deviations as the loss function to measure the difference between the network's predicted output and the actual expected output. The loss function is:

[0141] ,

[0142] Where N represents the number of adjustable potential samples. This represents the actual output of the i-th adjustable potential sampled data. This represents the predicted output of the i-th adjustable potential sampled data. This represents the l2 norm, used to calculate the square root of the sum of squares of vector elements;

[0143] Step 5: To compensate for the prediction bias of the base output, a margin function is introduced to construct a margin connection mechanism, the final prediction output of which is:

[0144] ,

[0145] in, For all sub-memory storage areas in the first The concatenated state vector is formed by concatenating the updated state vectors for each time period. This is a nonlinear residual function used to extract bias compensation information from the global state of the sub-memory storage area, and is ultimately used as the base output of the cascaded temporal memory network model with adjustable potential prediction. The superposition of these values ​​constitutes the predicted values ​​of the optimal adjustable potential prediction cascade temporal memory network model. ;

[0146] Step 6: Perform backpropagation based on the residual learning algorithm to solve for the gradient of the loss function with respect to each weight matrix; whereby, when updating the output weight matrix from the memory storage area to the time-domain output layer... At that time, according to the chain rule, the loss function right The gradient is:

[0147] ,

[0148] in, Indicates that the sub-memory storage area is in the first... The updated state vector for each time period; the loss function applied to the input weight matrix. and internal weight matrix To obtain the gradient, it is necessary to traverse all sub-memory storage areas and accumulate the gradient contribution of each storage area:

[0149] ,

[0150] ,

[0151] Step 7: Update the input weight matrix using the residual gradient iteration method. Internal weight matrix and output weight matrix ;

[0152] Determine if the current loss function has converged; if not, proceed to step 2; if yes, proceed to step 8.

[0153] Step 8: Output the global optimal parameter set The training of the prediction network is completed, resulting in the optimal adjustable potential prediction cascade temporal memory network model, and the globally optimal parameter set is obtained. Including the optimal input weight matrix Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector .

[0154] In one specific embodiment of the present invention, the residual gradient iteration method is used to update the input weight matrix. Internal weight matrix and output weight matrix The update formula used is:

[0155] ,

[0156] ,

[0157] ,

[0158] in, , , These are the updated input weight matrix, internal weight matrix, and output weight matrix, respectively. It is the learning rate, used to control the step size of each parameter update.

[0159] In one specific embodiment of the present invention, the adjustable potential prediction method further includes:

[0160] The stability of the optimal adjustable potential prediction cascaded temporal memory network model is determined using the following formula:

[0161] ,

[0162] in, This represents the internal weight matrix of the k-th sub-memory storage area optimized through margin learning. Indicates the spectral radius.

[0163] In one specific embodiment of the present invention, the acquired real-time adjustable potential data of distributed resources is input into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data:

[0164] Step 1: Obtain data related to the adjustable potential of distributed resources within the new sampling period, including the newly sampled adjustable power matrix. Electricity load matrix and the demand response signal matrix issued by the power grid side :

[0165] ,

[0166] in, For newly sampled adjustable potential data;

[0167] Step 2: Call the globally optimal parameter set obtained during the training phase, including the optimal input weight matrix. Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector The network parameters for the optimal adjustable potential prediction cascaded temporal memory network are set, and the initial state of each sub-memory storage area is set as the final state vector after convergence during the training phase. ;

[0168] Step 3: Newly sampled adjustable potential data The data is sequentially input into the optimal adjustable potential prediction cascaded temporal memory network according to the time series, and forward propagation calculation is performed. The state of each sub-memory storage area is updated in real time. For the k-th sub-memory storage area, the state update equation for time period t is:

[0169] ,

[0170] in, Indicates the t-th time period. The optimal state vector of each sub-memory storage area Indicates the first Time period The state vector of each sub-memory storage area This represents the input vector of the new adjustable potential sampling data at time t. Indicates the first Each sub-memory storage area for the first Cross-level feature coupling terms in individual memory storage areas;

[0171] Step 4: Concatenate the optimal state vectors of all sub-memory storage areas at time t to obtain the concatenated state vector. :

[0172] ,

[0173] In the formula, For sub-memory storage area The optimal state vector for time period t;

[0174] Based on the optimization results of the residual learning algorithm, the predicted value for time period t is calculated through the time-domain output layer:

[0175] ,

[0176] in, This represents the predicted value of the adjustable potential data of distributed resources on the electricity consumption side during time period t. Represents the state concatenation vector The diagonalized matrix;

[0177] Step 5: Repeat steps 3 and 4 to complete the adjustable potential prediction for all time periods within the new sampling period, and output the final adjustable potential prediction result matrix:

[0178] .

[0179] The adjustable potential prediction method in this invention will be described in detail below with reference to a specific implementation method.

[0180] like Figure 1 The diagram shown is a flowchart illustrating a method for constructing an adjustable potential prediction cascaded temporal memory network model according to an embodiment of the present invention. The specific steps are as follows:

[0181] Step A1: Construct the basic temporal memory network;

[0182] Step A2: Define the iteration count label t, and initialize it to 0;

[0183] Step A3: Initialize the single-layer basic temporal memory network. Initialize the number of iterations to 1. If the maximum number of iterations is not met, the number of iterations will be continuously increased until the output iteration count is met.

[0184] Step A4: Cascade the memory storage layers of multiple basic time-domain memory networks into cascaded sub-memory storage areas. Each memory storage layer serves as one of the sub-memory storage areas within the cascaded sub-memory storage area. Various types of adjustable potential sampling data are mapped to the input vectors of their respective sub-memory storage areas, with the input vector dimension matching the dimension of the input weight matrix of the corresponding sub-memory storage area. In specific implementation, when the adjustable potential data includes three types of data—adjustable power, electricity-side load value, and demand response issued by the grid side—the cascaded sub-memory storage area includes: Sub-memory storage area 1, used to process adjustable power, input... Sub-memory storage area 2 is used to process the load values ​​on the power consumption side and input them. Sub-memory storage area 3 is used to process demand responses issued by the power grid side, and input... The state update formulas for each basic temporal memory network are as follows:

[0185] ,

[0186] The formula for calculating its output data is:

[0187] ,

[0188] in, This represents the state vector of the memory storage layer at time t+1. This represents the state vector of the memory storage layer at time t. This represents the input of the feature input layer at time t. This represents the output of the temporal output layer of the basic temporal memory network at time t. This represents the bias vector. and Both represent activation functions;

[0189] Step A5: The cascaded sub-memory storage area comprises K sub-memory storage areas, and the cross-level coupling weight matrix between the k-th sub-memory storage area and the m-th sub-memory storage area is:

[0190] ,

[0191] The real-time state of the m-th sub-memory storage area is transferred to the k-th sub-memory storage area to achieve cross-level feature fusion. The state vector update equation of the k-th sub-memory storage area is:

[0192] ,

[0193] In the formula, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents a constant between 1 and 0, called the leakage rate, which controls how quickly the current state forgets past states; This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents the input weight matrix of the k-th basic temporal memory network feature input layer. This represents the input vector of the k-th basic temporal memory network feature input layer. Its function is to input the feature vector of the k-th basic temporal memory network into the input layer. Mapped to the neuron space of the k-th sub-memory storage area; This represents the internal weight matrix of the k-th sub-memory storage area, i.e., the connection weights between neurons within the same sub-memory storage area, used to transmit the historical state of that sub-memory storage area; This indicates that the m-th sub-memory storage area is in the... State vector for a given time period; coupling terms The m-th sub-memory storage area refers to the m-th sub-memory storage area in the... The state vector of a time period is coupled with a weight matrix across levels. After mapping, the feature interaction term is passed to the k-th sub-memory storage area to achieve deep fusion of data features at different levels; tanh is the activation function.

[0194] Where tanh is the activation function, and its formula is:

[0195] ,

[0196] Step A6: Train the adjustable potential prediction cascaded temporal memory network model using the residual learning algorithm to obtain the globally optimal parameter set; Step A7: Determine the stability of the optimized adjustable potential prediction cascaded temporal memory network model using the following formula:

[0197] ,

[0198] in, This represents the internal connection weight matrix of each sub-memory storage area optimized by margin learning. Indicates the spectral radius.

[0199] Step A8: Obtain the optimal adjustable potential prediction cascade temporal memory network model.

[0200] like Figure 2 As shown, a training flowchart for obtaining the optimal output weight matrix by training a network model using the residual learning algorithm is also provided. The specific steps are as follows:

[0201] Step B1: Initialize the relevant parameters of the residual learning algorithm and the key weight matrices of each layer of the cascaded temporal memory network. The number of sub-memory storage areas is... Each sub-memory storage area contains Q neurons, and the maximum number of iterations is T;

[0202] Step B2: Define the iteration count label t, and initialize it to 0;

[0203] Step B3: Initialize the number of iterations to 1. If the maximum number of iterations is not met, the number of iterations will continue to increase until the maximum number of iterations is met.

[0204] Step B4: Input the prepared historical adjustable potential data into the constructed cascaded temporal memory network model for forward propagation. For the k-th basic temporal memory network, update the memory storage area state according to the following formula:

[0205] ,

[0206] in, Indicates the first The state vector of the k-th sub-memory storage area in time period, This represents the input to the k-th sub-memory storage area in the t-th time period; This represents the state vector of the k-th sub-memory storage area in time period t;

[0207] Step B5: Calculate the output based on the state of the memory storage area:

[0208] ,

[0209] in, This represents the output of the temporal output layer of the k-th basic temporal memory network. This represents the output weight matrix of the k-th basic temporal memory network;

[0210] Ultimately, the basic output of the entire adjustable potential prediction cascaded temporal memory network model is... This is the set of outputs from the temporal output layers of each basic temporal memory network:

[0211] ,

[0212] in, This represents the output of the i-th sample in the temporal output layer of the k-th basic temporal memory network;

[0213] Step B6: Use the mean of squared deviations as the loss function to measure the difference between the network's predicted output and the actual expected output. The loss function is:

[0214] ,

[0215] Where N represents the number of samples, This represents the actual output of the i-th sample. This represents the predicted output for the i-th sample. This represents the l2 norm, used to calculate the square root of the sum of squares of vector elements;

[0216] Step B7: Introduce the residual learning module and the output of the residual connection. for:

[0217] ,

[0218] Among them, the surplus connection is applied in the training phase. For all sub-memory storage areas in the first The concatenated state vector is formed by concatenating the updated state vectors for each time period. This is a nonlinear residual function used to extract bias compensation information from the global state of the sub-memory storage area, and is ultimately used as the base output of the cascaded temporal memory network model with adjustable potential prediction. The superposition of these values ​​constitutes the predicted values ​​of the optimal adjustable potential prediction cascade temporal memory network model. ;

[0219] Step B8: Perform backpropagation using the residual learning algorithm to update the output weight matrix from the memory storage area to the temporal output layer. For example, according to the chain rule, the loss function right The gradient is:

[0220] ,

[0221] in, express Indicates that the sub-memory storage area is in the first... The updated state vector for each time period; the loss function applied to the input weight matrix. and internal weight matrix To obtain the gradient, it is necessary to traverse all sub-memory storage areas and accumulate the gradient contribution of each storage area:

[0222] ,

[0223] ,

[0224] Step B9: Update the weight matrix using the residual gradient iteration method. The update formula is as follows:

[0225] ,

[0226] ,

[0227] ,

[0228] in, It is the learning rate, which controls the step size for each parameter update;

[0229] Step B10: Determine whether the current loss function has converged; if not, proceed to step B3; if yes, proceed to step 11.

[0230] Step B11: Output the globally optimal parameter set The training of the network is completed, and the globally optimal parameter set is obtained. Including the optimal input weight matrix Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector .

[0231] like Figure 3 The diagram shown is a flowchart illustrating a method for predicting the adjustable potential of distributed resources on the power consumption side using an optimal adjustable potential prediction cascaded time-domain memory network model, as provided in an embodiment of the present invention. The specific process is as follows:

[0232] Step C1: Obtain data related to the adjustable potential of distributed resources within the new sampling period, including the newly sampled adjustable power matrix. Electricity load matrix and the demand response signal matrix issued by the power grid side :

[0233] ,

[0234] in, For the newly sampled adjustable potential data matrix;

[0235] Step C2: Initialize the parameters of the trained optimal adjustable potential prediction cascade temporal memory network by calling the globally optimal parameter set obtained during the training phase, including the optimal input weight matrix. Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix Optimal residual module weight matrix and optimal bias vector At the same time, the initial state of the memory storage area is set to the final state vector after convergence during the training phase. ;

[0236] Step C3: Newly sampled adjustable potential data The data is sequentially input into the optimal adjustable potential prediction cascaded temporal memory network according to the time series, and forward propagation calculation is performed to update the state of each sub-memory storage area in real time: For the k-th sub-memory storage, the state update equation for time period t is:

[0237] ,

[0238] in, Let represent the optimal state vector of the k-th sub-memory storage area in time period t. This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period This represents the input vector of the new adjustable potential sampling data in time period t. Indicates the first Each sub-memory storage area (lower-level sub-memory storage area) is related to the first... Cross-level feature coupling terms of individual sub-memory storage areas (high-level sub-memory storage areas);

[0239] Step C4: Vertically concatenate the optimal state vectors of all sub-memory storage areas at time t to obtain the concatenated state vector. :

[0240] ,

[0241] In the formula, Let be the optimal state vector of the k-th sub-memory storage area in time period t;

[0242] Based on the optimization results of the residual learning module, the predicted value for time period t is calculated through the time-domain output layer:

[0243] ,

[0244] in, This represents the predicted value of the adjustable potential data of distributed resources on the electricity consumption side during time period t. Represents the state concatenation vector The diagonalized matrix;

[0245] Step C5: Repeat steps C3 to C4 to complete the adjustable potential prediction for all time periods within the new sampling period, and output the final adjustable potential prediction result matrix:

[0246] .

[0247] Example 2

[0248] This invention provides an adjustable potential prediction device suitable for distributed resources on the electricity consumption side, comprising:

[0249] The historical adjustable potential data acquisition module is used to acquire historical adjustable potential data of distributed resources;

[0250] The module for obtaining the optimal adjustable potential prediction cascaded temporal memory network model is used to acquire a pre-constructed adjustable potential prediction cascaded temporal memory network model. Based on historical adjustable potential data, the module trains the adjustable potential prediction cascaded temporal memory network model using a residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model. The adjustable potential prediction cascaded temporal memory network model adopts a multi-sub-memory storage area cascaded stacked structure to achieve cross-level feature fusion.

[0251] The prediction module is used to input the real-time adjustable potential data of the acquired distributed resources into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data.

[0252] The rest are the same as in Example 1.

[0253] Example 3

[0254] This invention provides an adjustable potential prediction system for distributed resources on the power consumption side, including a storage medium and a processor;

[0255] The storage medium is used to store instructions;

[0256] The processor is configured to operate according to the instructions to execute the method according to any one of Embodiment 1.

[0257] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0258] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0259] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0260] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0261] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0262] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the adjustable potential of distributed resources on the electricity consumption side, characterized in that, include: Obtain historical adjustable potential data of distributed resources; A pre-constructed adjustable potential prediction cascaded temporal memory network model is obtained. Based on historical adjustable potential data, the adjustable potential prediction cascaded temporal memory network model is trained using the residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model. The adjustable potential prediction cascaded temporal memory network model adopts a multi-sub-memory storage area cascaded stacked structure to achieve cross-level feature fusion. The real-time adjustable potential data of the acquired distributed resources is input into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data.

2. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 1, characterized in that: The adjustable potential sampling data of the distributed resources includes adjustable power, load value on the electricity consumption side, and demand response issued by the grid side; The adjustable power is represented by a matrix. express: ， In the formula, This represents the actual adjustable power during time period t; The power-side load value is represented by a matrix. express: ， In the formula, This represents the total load on the electricity consumption side during time period t; The demand response issued by the power grid side express: ， In the formula, This represents the demand response value issued by the power grid side during time period t. This represents the device number of the demand response equipment issued by the power grid side during time period t.

3. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 1, characterized in that: The adjustable potential prediction cascaded temporal memory network model is constructed through the following steps: Step 1: Initialize the network parameters of the basic temporal memory network, which includes a feature input layer, a memory storage layer, and a temporal output layer arranged sequentially; the network parameters include the input weight matrix of the feature input layer. The internal weight matrix of the memory storage layer The output weight matrix of the time-domain output layer ,in, This indicates the input dimension of the feature input layer. This indicates the number of neurons in the memory storage layer. Indicates the output dimension of the time-domain output layer; Step 2: Cascade the memory storage layers of multiple basic temporal memory networks into cascaded sub-memory storage areas. Each memory storage layer serves as a sub-memory storage area within the cascaded sub-memory storage area. Various types of adjustable potential sampling data are mapped to the input vectors of the corresponding sub-memory storage areas, and the dimension of the input vectors matches the dimension of the input weight matrix of the corresponding sub-memory storage area. The cascaded sub-memory storage area comprises K sub-memory storage areas, and the cross-level coupling weight matrix between the k-th sub-memory storage area and the m-th sub-memory storage area is: ， The real-time state of the m-th sub-memory storage area is transferred to the k-th sub-memory storage area to achieve cross-level feature fusion. The state vector update equation of the k-th sub-memory storage area is: ， In the formula, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents a constant between 1 and 0, called the leakage rate, which controls how quickly the current state forgets past states; This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period; This represents the input weight matrix of the k-th basic temporal memory network feature input layer. This indicates that the k-th basic temporal memory network feature input layer is in the k-th... Input vector for a given time period Its function is to input the k-th basic temporal memory network feature into the layer at the th ... Input vector for a given time period Mapped to the neuron space of the k-th sub-memory storage area; This represents the internal weight matrix of the k-th sub-memory storage area, i.e., the connection weights between neurons within the same sub-memory storage area, used to transmit the historical state of that sub-memory storage area; This indicates that the m-th sub-memory storage area is in the... State vector for a given time period; coupling terms The m-th sub-memory storage area refers to the m-th sub-memory storage area in the... The state vector of a time period is coupled with a weight matrix across levels. The feature interaction term, after mapping, is passed to the k-th sub-memory storage area to achieve deep fusion of data features at different levels; tanh is the activation function. Step 3: Concatenate the time-domain output layers of multiple basic time-domain memory networks to obtain an adjustable potential prediction cascaded time-domain memory network model.

4. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 3, characterized in that: The state update formulas for each basic temporal memory network are as follows: ， The formula for calculating its output data is: ， in, This represents the state vector of the memory storage layer at time t+1. This represents the state vector of the memory storage layer at time t. This represents the input vector of the feature input layer of the basic temporal memory network at time t. This represents the output vector of the temporal output layer of the basic temporal memory network at time t. This represents the bias vector. and Both represent activation functions.

5. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 3, characterized in that: The process of training the adjustable potential prediction cascaded temporal memory network model using the residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model includes: Step 1: Initialize the relevant parameters of the residual learning algorithm and the weight matrices of each layer of the adjustable potential prediction cascaded temporal memory network model. The number of sub-memory storage areas is... Each sub-memory storage area contains Q neurons, and the maximum number of iterations is T; Step 2: Input the prepared historical adjustable potential data into the constructed cascaded temporal memory network model for forward propagation. For the k-th basic temporal memory network, update the memory storage area state according to the following formula: ， in, This indicates that the k-th sub-memory storage area is in the k-th sub-memory storage area. The state vector of a time period This indicates that the k-th basic temporal memory network feature input layer is in the k-th... The input vector for the time period; This represents the state vector of the k-th sub-memory storage area in time period t; Step 3: Based on the state vector of the k-th sub-memory storage area Calculate the output of the temporal output layer of the k-th basic temporal memory network: ， in, This represents the output of the temporal output layer of the k-th basic temporal memory network. This represents the output weight matrix of the k-th basic temporal memory network; Ultimately, the basic output of the entire adjustable potential prediction cascaded temporal memory network model is... This is the set of outputs from the temporal output layers of each basic temporal memory network: ， in, This represents the output of the temporal output layer of the i-th adjustable potential sampled data at the k-th basic temporal memory network; Step 4: Use the mean of squared deviations as the loss function to measure the difference between the network's predicted output and the actual expected output. The loss function is: ， Where N represents the number of adjustable potential samples. This represents the actual output of the i-th adjustable potential sampled data. This represents the predicted output of the i-th adjustable potential sampled data. This represents the l2 norm, used to calculate the square root of the sum of squares of vector elements; Step 5: To compensate for the prediction bias of the base output, a margin function is introduced to construct a margin connection mechanism, the final prediction output of which is: ， in, For all sub-memory storage areas in the first The concatenated state vector is formed by concatenating the updated state vectors for each time period. This is a nonlinear residual function used to extract bias compensation information from the global state of the sub-memory storage area, and is ultimately used as the base output of the cascaded temporal memory network model with adjustable potential prediction. The superposition of these values ​​constitutes the predicted values ​​of the optimal adjustable potential prediction cascade temporal memory network model. ; Step 6: Perform backpropagation based on the residual learning algorithm to solve for the gradient of the loss function with respect to each weight matrix; whereby, when updating the output weight matrix from the memory storage area to the time-domain output layer... At that time, according to the chain rule, the loss function right The gradient is: ， in, Indicates that the sub-memory storage area is in the first... The updated state vector for each time period; the loss function applied to the input weight matrix. and internal weight matrix To obtain the gradient, it is necessary to traverse all sub-memory storage areas and accumulate the gradient contribution of each storage area: ， ， Step 7: Update the input weight matrix using the residual gradient iteration method. Internal weight matrix and output weight matrix ; Determine if the current loss function has converged; if not, proceed to step 2; if yes, proceed to step 8. Step 8: Output the global optimal parameter set The training of the prediction network is completed, resulting in the optimal adjustable potential prediction cascade temporal memory network model, and the globally optimal parameter set is obtained. Including the optimal input weight matrix Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector .

6. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 5, characterized in that: Update the input weight matrix using the residual gradient iteration method. Internal weight matrix and output weight matrix The update formula used is: ， ， ， in, , , These are the updated input weight matrix, internal weight matrix, and output weight matrix, respectively. It is the learning rate, used to control the step size of each parameter update.

7. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 6, characterized in that: The adjustable potential prediction method further includes: The stability of the adjustable potential prediction cascade temporal memory network model optimized by residual learning is determined using the following formula: ， in, This represents the internal weight matrix of the k-th sub-memory storage area optimized through margin learning. Indicates the spectral radius.

8. The adjustable potential prediction method for distributed resources on the power consumption side according to claim 5, characterized in that: The acquired real-time adjustable potential data of distributed resources is input into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data: Step 1: Obtain data related to the adjustable potential of distributed resources within the new sampling period, including the newly sampled adjustable power matrix. Electricity load matrix and the demand response signal matrix issued by the power grid side : ， in, For newly sampled adjustable potential data; Step 2: Call the globally optimal parameter set obtained during the training phase, including the optimal input weight matrix. Optimal internal weight matrix Optimal cross-level coupling weight matrix Optimal output weight matrix The output weight matrix optimized through residual learning and optimal bias vector The network parameters for the optimal adjustable potential prediction cascaded temporal memory network are set, and the initial state of each sub-memory storage area is set as the final state vector after convergence during the training phase. ; Step 3: Newly sampled adjustable potential data The data is sequentially input into the optimal adjustable potential prediction cascaded temporal memory network according to the time series, and forward propagation calculation is performed. The state of each sub-memory storage area is updated in real time. For the k-th sub-memory storage area, the state update equation for time period t is: ， in, Indicates the first The optimal state vector of each sub-memory storage area in time period t. Indicates the first The sub-memory storage area is in the first The state vector of a time period This represents the input vector of the new adjustable potential sampling data in time period t. Indicates the first Each sub-memory storage area for the first Cross-level feature coupling terms in individual memory storage areas; Step 4: Concatenate the optimal state vectors of all sub-memory storage areas at time t to obtain the concatenated state vector. : ， In the formula, For the first The optimal state vector of each sub-memory storage area in time period t; Based on the optimization results of the residual learning algorithm, the predicted value for time period t is calculated through the time-domain output layer: ， in, This represents the predicted value of the adjustable potential data of distributed resources on the electricity consumption side during time period t. Represents the state concatenation vector The diagonalized matrix; Step 5: Repeat steps 3 and 4 to complete the adjustable potential prediction for all time periods within the new sampling period, and output the final adjustable potential prediction result matrix: 。 9. A device for predicting the adjustable potential of distributed resources on the power consumption side, characterized in that, include: The historical adjustable potential data acquisition module is used to acquire historical adjustable potential data of distributed resources; The module for obtaining the optimal adjustable potential prediction cascaded temporal memory network model is used to acquire a pre-constructed adjustable potential prediction cascaded temporal memory network model. Based on historical adjustable potential data, the module trains the adjustable potential prediction cascaded temporal memory network model using a residual learning algorithm to obtain the optimal adjustable potential prediction cascaded temporal memory network model. The adjustable potential prediction cascaded temporal memory network model adopts a multi-sub-memory storage area cascaded stacked structure to achieve cross-level feature fusion. The prediction module is used to input the real-time adjustable potential data of the acquired distributed resources into the optimal adjustable potential prediction cascaded time-domain memory network model to obtain the predicted value of the adjustable potential data.

10. A system for predicting the adjustable potential of distributed resources on the electricity consumption side, characterized in that: Including storage media and processor; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the method according to any one of claims 1-8.

Images