Transformer area load regulation and control prediction method based on multi-gating hybrid long short-term memory network

Abstract

The invention relates to a transformer area load regulation and control prediction method based on a multi-gating hybrid long short-term memory network. The method comprises the following steps: S1, obtaining historical power load data and influence factor characteristics which need to be considered and have influences on power loads; s2, constructing an input / output data set model; s3, constructing a core network composed of a plurality of LSTMs, wherein the core network is used for extracting periodic and uncertain feature data; s4, constructing a multi-gating unit network associated with the output task, and performing fusion output on the multi-dimensional features to obtain a load component prediction value; and S5, constructing and training a multi-gating hybrid long-short-term memory network prediction model, fusing different frequency load components, and adding a Huber normal form into a quantile loss function to correct a front-end model error to obtain a final power load prediction result. According to the invention, the power distribution area power load flexible resource credible regulation and control prediction precision can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of control and prediction of power load in distribution station area, and relates to a reliable control and prediction method of flexible resources of power load in distribution station area, in particular to a load control and prediction method of distribution station area based on multi-gated hybrid long short-term memory network. Background Art

[0002] The distribution network is an important part of the power system. It receives electricity from the power supply side and distributes it to various users step by step or on-site through distribution facilities. Distribution equipment is the last mile of power supply for power companies to serve users. It will directly supply power to various users such as residents, enterprises, industry and commerce, and schools. There are many devices, a wide distribution range, and a complex surrounding environment. It is easily affected by various external environments and construction. The increasing number of power equipment and facilities has caused great difficulties in traditional operation and maintenance management. With the access of a large number of new elements such as distributed photovoltaics, energy storage, electric vehicles, and electric heating to distribution stations, it not only brings challenges to the safe operation of medium and low voltage distribution networks, but also contains huge potential for flexible adjustment.

[0003] Low-voltage power grids in most provinces and cities in my country adopt radial design and operation mode. The advantages are simple operation mode of power grid and easy implementation of power management; the disadvantages are lack of operational flexibility, inability to realize load transfer between power grids and optimal distribution of load in power grids. However, the flexibility resources of a single power grid are limited, unable to meet the load peak and reactive voltage regulation of the power grid, which has a great impact on the reliability of power supply in the power grid. In urban distribution networks, due to the limited power supply capacity at the end of the distribution network, accompanied by a sharp increase in load, it presents a regional intensive and unbalanced state. Load-intensive areas such as business districts, old urban areas, and tourist attractions are limited by geographical location, and the capacity of line corridors is insufficient; in rural distribution networks, since rural loads are mainly composed of wells, irrigation, small industries, and tourism, they show strong seasonal and cyclical fluctuations, and some power grids are overloaded and overloaded.

[0004] Although extensive research has been conducted on the technology of integrated terminal in substations at home and abroad, the existing research either focuses only on the edge communication technology of integrated terminal in substations; or focuses on the status detection of distribution substations; or the substations consider the use of flexible resources, etc., but no accurate evaluation of the adjustable capacity of distribution substations, the mutual influence of flexible resources, the key technologies of coordinated optimization of multi-substation flexible resources driven by integrated terminal data, and the value of integrated terminal data have been conducted.

[0005] Therefore, research on the adjustable capacity of flexible resources in distribution substations and the reliable regulation and prediction technology for optimizing flexible resources is of great significance to the low-carbon operation of power systems.

[0006] After searching, no public documents of the prior art that are identical or similar to the present invention were found. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a load control prediction method for a substation based on a multi-gated hybrid long short-term memory network, which can solve the technical problem of studying the correction method of integrating terminal spatial aggregation data and time decoupled data under different control scenarios to obtain accurate online calculation and predictive evaluation of trusted control capacity.

[0008] The present invention solves the practical problem by adopting the following technical solutions:

[0009] A load control prediction method for a substation based on a multi-gated hybrid long short-term memory network includes the following steps:

[0010] S1. Obtain historical power load data and the characteristics of factors affecting power load that need to be considered, and pre-process the data;

[0011] S2, constructing an input / output data set model based on the data preprocessed in step S1;

[0012] S3, based on the input data set provided in step S2, construct a core network composed of multiple LSTMs to extract periodicity and uncertainty feature data;

[0013] S4, based on the periodicity and uncertainty feature data extracted in step S3, construct a multi-gated unit network associated with the output task, fuse and output the multi-dimensional features, and obtain the load component prediction value;

[0014] S5. Input the load component prediction value obtained in step S4 into the multi-gated hybrid long short-term memory network, construct and train the multi-gated hybrid long short-term memory network prediction model, integrate the load components of different frequencies, and use the Huber paradigm added to the quantile loss function to correct the front-end model error to obtain the final power load prediction result.

[0015] Moreover, the characteristics of the influencing factors affecting the power load in the distribution station area considered in step S1 include: meteorological characteristics and social factors; the data preprocessing method in step S1 includes data cleaning, outlier processing, and data normalization.

[0016] Moreover, the specific method of step S2 is:

[0017] The time series length of the input / output data is set to 48 hours; among them, the input data set model features select the total load of the region, the total load of each area in the region, the date and hour code, and the holiday code.

[0018] Moreover, the specific method of step S3 is:

[0019] The gate structure in the LSTM network filters the received information through opening and closing operations. The basis for the screening is the weight set of each gate, and the weight value will change with the training of the LSTM neural network. When the model is actually applied, each gate can use the obtained weight parameter set to process its own input. After the information is processed by all gate structures, the predicted value at the current moment can be output;

[0020] Among them, the training process of the LSTM network is divided into two parts: forward propagation and back propagation. The information passes through the forget gate, input gate, and output gate in sequence, and the final output result is the forward propagation process. The back propagation process of LSTM is to iteratively update the parameters through the gradient descent method. Through multiple continuous training of LSTM, the periodicity and uncertainty feature data are finally extracted.

[0021] Moreover, the specific method of step S4 is:

[0022] (1) The number of gated units and tasks must be consistent to ensure that the mapping between tasks and features is unique. Considering that the fully connected layer is essentially a weight stacking layer, the Add layer in the multi-gated mixture of experts network (MMoE) is discarded, and the feature fusion method is changed from vertical splicing to horizontal splicing.

[0023] (2) Based on the multi-task learning machine with soft parameter sharing mechanism, multiple parallel networks are constructed to mine features from multiple dimensions. At the same time, the parameters will tend to be similar during the training process to ensure the unity of prediction purpose between networks. The multi-dimensional features are fused and output to obtain the load component prediction value.

[0024] Moreover, the specific method of step S5 is:

[0025] The load component prediction value obtained in step S4 is input into a multi-gated hybrid long short-term memory network and trained. The load components of different frequencies are integrated. When constructing an accurate prediction model, MSE or MAE is often used as the Loss function and combined with the Adam optimization algorithm to find the optimal solution of the model. The Huber paradigm is added to the quantile loss function to correct the front-end model error. According to the output data set structure in model S2, the final power load prediction result is finally obtained.

[0026] Advantages and beneficial effects of the present invention:

[0027] 1. The present invention proposes a load control prediction method for a substation based on a multi-gated hybrid long-short-term memory network. First, the historical power load data of the distribution substation area and the characteristics of related influencing factors are obtained, and the data is preprocessed; secondly, the input / output data set model is constructed based on the total load of the area and the date, hour code and holiday code. Then, a core network composed of multiple LSTMs is constructed to extract periodicity and uncertainty characteristics, and a multi-gated unit network associated with the output task is used to fuse and output multidimensional features; finally, a multi-gated hybrid long-short-term memory network prediction model is constructed to fuse different frequency load components, and the Huber paradigm is added to the quantile loss function to correct the front-end model error to obtain the final power load prediction result. The present invention interconnects and supplies between substations, coordinates the dispatch of flexible resources in the substation area, alleviates the load peak of a single substation area, balances the load between substations, and improves the utilization rate of flexible resources. The present invention conducts research on coordinated management and control of flexible resources in substations, which can deeply explore and utilize the flexibility of multiple heterogeneous resources scattered in substations, divide and coordinate the scheduling of small-scale adjustable resources, and provide a large amount of flexibility support for the low-carbon, economical and safe operation of new distribution systems, which is of great significance to the realization of my country's "dual carbon" goals.

[0028] 2. The present invention proposes a multi-gated hybrid long short-term memory network model, which is divided into two parts. One is a core network composed of multiple LSTMs for extracting periodic and uncertainty features; the other is a gating unit network associated with the output task to fuse and output multidimensional features. The number n of LSTMs in the core network layer has no actual correlation with the number of tasks, but when n is greater than or equal to the number of tasks, the efficiency advantage of the multi-task model compared to the traditional multi-step prediction method is lost. At the same time, the number of gating units must be consistent with the number of tasks, thereby ensuring that the mapping between tasks and features is unique. Considering that the fully connected layer is essentially a weight superposition layer, the Add layer in the MMoE is discarded, and the feature fusion method is changed from vertical splicing to horizontal splicing, thereby improving the prediction accuracy of the reliable regulation of flexible resources for power load in the distribution station area.

[0029] 3. The present invention proposes a multi-gated hybrid long short-term memory network model, adopts the quantile loss function for model training, adds the Huber paradigm to the quantile loss function to make the quantile loss function differentiable everywhere, and substitutes it into the Lossq function to find the optimal solution of the model, thereby correcting the front-end model error and improving the prediction accuracy of the reliable regulation of flexible resources of power load in the distribution station area. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flow chart of a method for predicting load control in a substation based on a multi-gated hybrid long short-term memory network provided by an embodiment of the present invention;

[0031] Figure 2A diagram of the structure of a long short-term memory neural network (LSTM) unit provided in an embodiment of the present invention;

[0032] Figure 3 A framework diagram of a multi-gate mixture-of-experts (MMoE) model provided in an embodiment of the present invention;

[0033] Figure 4 An input / output structure diagram of a multi-gated hybrid long short-term memory network provided in an embodiment of the present invention;

[0034] Figure 5 A structural diagram of a multi-gate mixture-of-LSTMs (MMoL) model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0036] A load control prediction method for a substation based on a multi-gated hybrid long short-term memory network includes the following steps:

[0037] S1. Obtain historical power load data and the characteristics of factors affecting power load that need to be considered, and pre-process the data;

[0038] The characteristics of the influencing factors affecting the power load in the distribution station area considered in step S1 include: meteorological characteristics (temperature, humidity, air pressure, dew point) and social factors (holiday characteristics, monthly characteristics, weekly characteristics).

[0039] The data preprocessing method in step S1 includes data cleaning, outlier processing, and data normalization.

[0040] S2, constructing an input / output data set model based on the data preprocessed in step S1;

[0041] The specific method of step S2 is:

[0042] The time series length of input / output data is set to 48 hours. The input data set model features include the total load of the region, the total load of each area in the region, the date and hour code, and the holiday code.

[0043] In this embodiment, the input / output data set model is constructed in step S2, and the time series length of the input / output data setting is 48 hours.

[0044] Reasonable setting of input data set can fundamentally reduce the amount of model calculation and help the model extract effective features for learning. By constructing the output data set, the multi-step prediction problem is transformed into a time series multi-task prediction problem, which helps the model to clarify the output task.

[0045] In this embodiment, the input / output data set model is constructed in step S2, and the output length is that the model of this project predicts the load per hour within a short period of 48 hours, so the output is the next 48 hours at the prediction time. The input data composition needs to consider the input feature quantity and time series length. Too many feature quantities and too long time series length will affect the model prediction efficiency. Therefore, the input data set features select the total load of the region, the total load of each area in the region, the date and hour code, and the holiday code. Among them, the holiday code, the code value on the holiday day is 1, and the working day is 0. The input data set time series length is also selected as 48h, which ensures the prediction efficiency while mining the feature information as much as possible.

[0046] S3, based on the input data set provided in step S2, construct a core network composed of multiple LSTMs to extract periodicity and uncertainty feature data;

[0047] The specific method of step S3 is:

[0048] The gate structure in the LSTM network filters the received information through opening and closing operations. The basis for the screening is the weight set of each gate, and the weight value will change with the training of the LSTM neural network. When the model is actually applied, each gate can use the obtained weight parameter set to process its own input. After the information is processed by all gate structures, the predicted value at the current moment can be output;

[0049] Among them, the training process of the LSTM network is divided into two parts: forward propagation and back propagation. The information passes through the forget gate, input gate, and output gate in sequence, and the final output result is the forward propagation process. The back propagation process of LSTM is to iteratively update the parameters through the gradient descent method. Through multiple continuous training of LSTM, the periodicity and uncertainty feature data are finally extracted.

[0050] In this embodiment, the core concepts in the LSTM network are cell states and various gate structures. The cell state stores the key information of the sequence, allowing the information to be passed along the sequence chain, which is equivalent to the "memory" of the LSTM network. The gate structure can filter the received information through the operation of opening and closing, and the basis for the screening is the weight set of each gate, and the weight value will change with the training of the LSTM neural network. When the model is actually applied, each gate can use the obtained weight parameter set to process its own input. After the information is processed by all gate structures, the predicted value of the current moment (missing moment) can be output. The training process of LSTM is divided into two parts: forward propagation and back propagation. Information passes through the forget gate, input gate, and output gate in turn, and the output result is finally obtained, which is the forward propagation process. The back propagation process of LSTM is to iteratively update the parameters through the gradient descent method. Through the continuous training of multiple LSTMs, the periodicity and uncertainty feature data are finally extracted.

[0051] S4, based on the periodicity and uncertainty feature data extracted in step S3, construct a multi-gated unit network associated with the output task, fuse and output the multi-dimensional features, and obtain the load component prediction value;

[0052] like Figure 3 As shown, the specific method of step S4 is:

[0053] (1) The number of gated units and tasks must be consistent to ensure that the mapping between tasks and features is unique. Considering that the fully connected layer is essentially a weight stacking layer, the Add layer in the multi-gated mixture of experts network (MMoE) is discarded, and the feature fusion method is changed from vertical splicing to horizontal splicing.

[0054] (2) Based on the multi-task learning machine with soft parameter sharing mechanism, multiple parallel networks are constructed to mine features from multiple dimensions. At the same time, the parameters will tend to be similar during the training process to ensure the unity of prediction purpose between networks. The multi-dimensional features are fused and output to obtain the load component prediction value.

[0055] S5. Input the load component prediction value obtained in step S4 into the multi-gated hybrid long short-term memory network, construct and train the multi-gated hybrid long short-term memory network prediction model, integrate the load components of different frequencies, and use the Huber paradigm added to the quantile loss function to correct the front-end model error to obtain the final power load prediction result.

[0056] like Figure 5 As shown, the specific method of step S5 is:

[0057] The load component prediction value obtained in step S4 is input into a multi-gated hybrid long short-term memory network and trained. The load components of different frequencies are integrated. When constructing an accurate prediction model, MSE or MAE is often used as the Loss function and combined with the Adam optimization algorithm to find the optimal solution of the model. The Huber paradigm is added to the quantile loss function to correct the front-end model error. According to the output data set structure in model S2, the final power load prediction result is finally obtained.

[0058] The present invention is further described below by means of specific examples:

[0059] Embodiment 1:

[0060] like Figure 1-5 As shown, the load control prediction method for the distribution area based on the multi-gated hybrid long short-term memory network provided in this embodiment, the distribution area power load flexible resource trusted control prediction method includes the following steps:

[0061] S1. Obtain historical power load data and the characteristics of factors affecting the power load that need to be considered, and pre-process the data.

[0062] In this embodiment, the characteristics of the influencing factors affecting the power load considered in the historical power load data mainly include meteorological characteristics (temperature, humidity, air pressure, dew point) and social factors (holiday characteristics, monthly characteristics, weekly characteristics).

[0063] Data preprocessing methods include data cleaning, outlier processing, and data normalization. Data cleaning is to fill in missing data values ​​using upward filling, downward filling, and daily load average filling methods; outlier processing refers to removing typical outliers in the data, such as extremely large load data, negative numbers, and other load data that are impossible to appear; data normalization refers to normalizing the load data using the maximum and minimum value normalization method, and the calculation formula (1) is:

[0064]

[0065] In the formula, x represents the original vector, x min and x max are the minimum and maximum values ​​of the samples respectively, and x' is the normalized vector.

[0066] S2. Build input / output dataset model

[0067] The time series length of the input / output data settings is 48 hours. The input data set features include the total load of the region, the total load of each area in the region, the date and hour code, and the holiday code.

[0068] Reasonable setting of input data set can fundamentally reduce the amount of model calculation and help the model extract effective features for learning. By constructing output data set, multi-step prediction problem is converted into time series multi-task prediction problem, which helps the model to clarify output task. Figure 4 As shown. The time series length of the input / output data settings is 48 hours. The output length is that the model of the present invention predicts the load per hour within a short period of 48 hours, so the output is the next 48 hours at the prediction time. The composition of the input data needs to consider the input feature quantity and the time series length. Too many feature quantities and too long time series length will affect the prediction efficiency of the model. Therefore, the input data set features select the total load of the region, the total load of each area in the region, the date and hour code, and the holiday code. Among them, the holiday code, the code value on the holiday day is 1, and the working day is 0. The time series length of the input data set is also selected as 48h, which ensures the prediction efficiency while mining the feature information as much as possible.

[0069] S3. Build a core network model composed of multiple LSTMs

[0070] Relying on the input data set provided by S2, a core network composed of multiple LSTMs is constructed to extract periodic and uncertainty features; the number n of LSTMs in the core network layer has no actual correlation with the number of tasks, but when n is greater than or equal to the number of tasks, the efficiency advantage of the multi-task model over the traditional multi-step prediction method is lost.

[0071] Long-Short Memory Network (LSTM) is often used to solve time series prediction problems. LSTM neurons have three logic control units: forget gate layer, input gate layer, and output gate layer, which enable LSTM to remember the periodic characteristics of data well. The internal structure of LSTM nodes is as follows: Figure 2 As shown in the figure. LSTM also inherits the advantages of neural networks, that is, mining the nonlinear characteristics of uncertain load fluctuations in load data. However, a single LSTM cannot fully mine all the features hidden in the data, nor can it meet the prediction requirements of multi-step prediction.

[0072] S4. Constructing a multi-gated unit network

[0073] The extracted periodicity and uncertainty feature data obtained in step S3 are used to construct a multi-gated unit network associated with the output task, and the multi-dimensional features are fused and output. The number of gated units must be consistent with the number of tasks to ensure that the mapping between tasks and features is unique. Considering that the fully connected layer is essentially a weight superposition layer, the patent of this project discards the Add layer in the multi-gate mixture-of-experts (MMoE) network. The feature fusion method is also changed from vertical splicing to horizontal splicing.

[0074] The multi-gate mixture of experts (MMoE) is a multi-task learning machine based on a soft parameter sharing mechanism. Compared with the hard parameter sharing multi-task learning machine, it can construct multiple parallel networks to mine features from multiple dimensions. At the same time, the parameters will tend to be similar during the training process to ensure the unity of the prediction purpose between networks. The input data first passes through the network layer composed of multiple expert networks. The expert network uses the Relu nonlinear function as the activation function neural network, and its formula is shown in formula (2):

[0075] Expert(X) i =RelU(ω i X+bias i ) (2)

[0076] Where i represents the i-th expert network; X is a matrix composed of n inputs x; ω is the Expert initialization weight matrix; bias is the bias matrix.

[0077] The output features of multiple expert networks need to be screened and fused by gating units with the same number of tasks. The specific formula of the gating unit is shown in formula (3). Finally, the task output is obtained through the tower layer formula (4):

[0078]

[0079] y k =h k (Gate(x) k ) (4)

[0080] Where k represents the kth task. From the above process, we can see that the expert network in MMoE is the core network of the model. Due to the use of nonlinear activation functions, the expert network has significant advantages in image or spatial feature extraction, but it is difficult to effectively mine the temporal features for time series problems. At the same time, the multi-gating mechanism proposed by MMoE reasonably solves the mapping relationship between the features mined by the core network and each output task.

[0081] S5. Constructing a multi-gated hybrid long short-term memory network prediction model

[0082] In step S4, the load component prediction value is obtained and input into the multi-gated hybrid long short-term memory network and trained, and the load components of different frequencies are integrated. The Huber paradigm is added to the quantile loss function to correct the front-end model error, and the final power load prediction result is obtained.

[0083] When building an accurate prediction model, MSE or MAE is often used as the Loss function and combined with the Adam optimization algorithm to find the optimal solution of the model. This method can only obtain the specific expected value of the future load. The probability prediction combines a specific function in the model, and the model can output the load prediction probability interval. Considering that accurate prediction and probability prediction results can provide load prediction information from different angles, the quantile loss function is used for model training, and the function expression is shown in formula (5). Since the quantile loss function is not differentiable at the origin, it is not suitable for the Adam optimization algorithm. Therefore, this patent adopts the quantile loss function to add the Huber paradigm to make the quantile loss function differentiable everywhere. The Huber paradigm function expression is shown in formula (6). Substituting it into the Lossq function, Losshq can be obtained as shown in formula (7).

[0084]

[0085] Where q is the quantile, and the value range is [0.1, 0.9]; is the predicted value of the model under the current quantile value; y t is the true value of the load; ε is the Huber normal threshold, which is the only hyperparameter of the Loss function.

[0086] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0087] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0088] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

Claims

1. A load control prediction method for a distribution area based on a multi-gated hybrid long short-term memory network, characterized in that: The steps include: S1. Obtain historical power load data and the characteristics of factors affecting power load that need to be considered, and pre-process the data; S2, constructing an input / output data set model based on the data preprocessed in step S1; S3, based on the input data set provided in step S2, construct a core network composed of multiple LSTMs to extract periodicity and uncertainty feature data; S4, based on the periodicity and uncertainty feature data extracted in step S3, construct a multi-gated unit network associated with the output task, fuse and output the multi-dimensional features, and obtain the load component prediction value; S5. Input the load component prediction value obtained in step S4 into the multi-gated hybrid long short-term memory network, construct and train the multi-gated hybrid long short-term memory network prediction model, integrate the load components of different frequencies, and use the Huber paradigm added to the quantile loss function to correct the front-end model error to obtain the final power load prediction result.

2. According to claim 1, a method for predicting load control in a substation based on a multi-gated hybrid long short-term memory network is characterized in that: The characteristics of the influencing factors affecting the power load in the distribution station area considered in step S1 include: meteorological characteristics and social factors; the data preprocessing method in step S1 includes data cleaning, outlier processing, and data normalization.

3. According to claim 1, a method for predicting load control in a substation based on a multi-gated hybrid long short-term memory network is characterized in that: The specific method of step S2 is: The time series length of the input / output data is set to 48 hours; among them, the input data set model features select the total load of the region, the total load of each area in the region, the date and hour code, and the holiday code.

4. According to claim 1, a method for predicting load control in a substation based on a multi-gated hybrid long short-term memory network is characterized in that: The specific method of step S3 is: The gate structure in the LSTM network filters the received information through opening and closing operations. The basis for the screening is the weight set of each gate, and the weight value will change with the training of the LSTM neural network. When the model is actually applied, each gate can use the obtained weight parameter set to process its own input. After the information is processed by all gate structures, the predicted value at the current moment can be output; Among them, the training process of the LSTM network is divided into two parts: forward propagation and back propagation. The information passes through the forget gate, input gate, and output gate in sequence, and the final output result is the forward propagation process. The back propagation process of LSTM is to iteratively update the parameters through the gradient descent method. Through multiple continuous training of LSTM, the periodicity and uncertainty feature data are finally extracted.

5. The method for predicting load control in a substation based on a multi-gated hybrid long short-term memory network according to claim 1 is characterized in that: The specific method of step S4 is: (1) The number of gated units and tasks must be consistent to ensure that the mapping between tasks and features is unique. Considering that the fully connected layer is essentially a weight stacking layer, the Add layer in the multi-gated mixture of experts network (MMoE) is discarded, and the feature fusion method is changed from vertical splicing to horizontal splicing. (2) Based on the multi-task learning machine with soft parameter sharing mechanism, multiple parallel networks are constructed to mine features from multiple dimensions. At the same time, the parameters will tend to be similar during the training process, ensuring the unity of prediction purpose between networks; The multi-dimensional features are fused and output to obtain the load component prediction value.

6. The method for predicting load control in a substation based on a multi-gated hybrid long short-term memory network according to claim 1 is characterized in that: The specific method of step S5 is: The load component prediction value obtained in step S4 is input into a multi-gated hybrid long short-term memory network and trained. The load components of different frequencies are integrated. When constructing an accurate prediction model, MSE or MAE is often used as the Loss function and combined with the Adam optimization algorithm to find the optimal solution of the model. The Huber paradigm is added to the quantile loss function to correct the front-end model error. According to the output data set structure in model S2, the final power load prediction result is finally obtained.

Images