A method and system for dynamically regulating the load of a distributed new energy grid connection

Abstract

The present invention relates to the technical field of power grid load regulation, and discloses a method and system for dynamically regulating the load of a distributed new energy grid connection, including obtaining multi-source data of the power grid, dynamically aligning and feature extracting the multi-source data to obtain a multi-source feature tensor, inputting the multi-source feature tensor into a spatio-temporal encoding model to obtain a spatio-temporal encoding matrix; inputting the spatio-temporal encoding matrix into a load prediction model to obtain a load prediction result; obtaining the real-time operation data of the power grid, and dynamically regulating the load of the power grid according to the real-time operation data and the load prediction result. The present invention solves the problem of inconsistent time granularity through multi-source data alignment optimization, effectively separates short-term fluctuations and long-term trends through a spatio-temporal attention template and a spatio-temporal encoding model, and improves the accuracy of the load prediction result through cross-scale joint modeling. The present invention can improve the efficiency of dispatching response and provide high-timeliness support for the intelligent dispatching of a new energy high-penetration power grid.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid load regulation, and particularly to a method and system for dynamically regulating the load of a power grid with distributed new energy access. Background Art

[0002] With the large-scale grid connection of distributed new energy such as wind power and photovoltaic power, the power grid faces many challenges in load regulation operation. On the one hand, the data sources obtained by the power grid are diverse and the time granularities are different. For example, new energy power generation data may be collected at a high frequency in minutes to reflect the instantaneous fluctuations of photovoltaic or wind power, while meteorological data is often recorded at a cycle of hours or days to record the slowly changing characteristics such as temperature and wind speed, and load data may be collected at fixed intervals, but there are sampling missing or anomalies due to holidays or emergencies. Due to the heterogeneity of multi-source data, it is difficult for the power grid to effectively align and fuse multi-source data when performing load regulation analysis; on the other hand, due to the short-term fluctuations (such as photovoltaic cloud occlusion) and long-term trends (such as meteorological changes) of new energy output due to meteorological factors, this short-term and short-term coupling effect leads to insufficient accuracy of the model results of using a single model for data analysis and prediction; in addition, most of the current power grid regulations rely on offline optimization and rule bases, resulting in delays in command generation, leading to delayed dispatching responses and difficulty in coping with the rapid fluctuations of new energy.

[0003] In view of the problems of multi-source data heterogeneity, multi-time scale feature conflicts and dispatching response lags existing in the above-mentioned distributed new energy access to the power grid, there is an urgent need for a load dynamic regulation method suitable for the operation optimization of an intelligent power grid with high proportion of energy penetration. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a method and system for dynamically regulating the load of a power grid with distributed new energy access, so as to solve the problems of multi-source data heterogeneity, multi-time scale feature conflicts and dispatching response lags existing in the existing distributed new energy access to the power grid, and achieve the effect of improving the accuracy of power grid load assessment and the real-time response ability of power grid dispatching.

[0005] In a first aspect, the present invention provides a method for dynamically regulating the load of a power grid with distributed new energy access, the method comprising:

[0006] Obtain multi-source data of distributed new energy access to the power grid, and perform dynamic alignment and feature extraction on the multi-source data to obtain a multi-source feature tensor, where the multi-source data includes new energy power generation data, meteorological data and load data;

[0007] Input the multi-source feature tensor into a preset spatio-temporal encoding model to obtain a spatio-temporal encoding matrix, where the spatio-temporal encoding model is constructed based on a deep learning model with a multi-head attention mechanism;

[0008] Input the spatio-temporal coding matrix into a preset load forecasting model to obtain a load forecasting result, where the load forecasting model is obtained based on cross-scale joint modeling of a gated recurrent unit and a long short-term memory neural network;

[0009] Obtain the real-time operation data of distributed new energy accessing the power grid, and dynamically regulate the load of distributed new energy accessing the power grid according to the real-time operation data and the load forecasting result.

[0010] Further, before the step of inputting the multi-source feature tensor into a preset spatio-temporal coding model to obtain a spatio-temporal coding matrix, it further includes:

[0011] Perform velocity stratification feature engineering on the multi-source feature tensor to obtain a velocity stratification feature matrix;

[0012] Input the velocity stratification feature matrix into a preset weight mapping model to obtain a spatio-temporal attention template, and embed the spatio-temporal attention template into the preset spatio-temporal coding model, where the weight mapping model is constructed based on a multi-layer perceptron.

[0013] Further, the step of performing velocity stratification feature engineering on the multi-source feature tensor to obtain a velocity stratification feature matrix includes:

[0014] Calculate the fluctuation variance of the new energy power generation data, the trend sensitivity of the meteorological data, and the stationarity index of the load data respectively;

[0015] Combine the fluctuation variance, the trend sensitivity, and the stationarity index to obtain a velocity stratification feature matrix.

[0016] Further, the step of inputting the velocity stratification feature matrix into a preset weight mapping model to obtain a spatio-temporal attention template includes:

[0017] Map the velocity stratification feature matrix into attention weights through a weight mapping model, where the attention weights include short-term attention weights and long-term attention weights;

[0018] Smoothly optimize the attention weights through a dynamic programming algorithm to generate a spatio-temporal attention template.

[0019] Further, the step of inputting the multi-source feature tensor into a preset spatio-temporal coding model to obtain a spatio-temporal coding matrix includes:

[0020] According to the multi-head attention separation mechanism of the spatio-temporal coding model, perform feature splitting on the multi-source feature tensor;

[0021] Enhance spatio-temporal position encoding according to dynamic load information to obtain a spatio-temporal encoding matrix, where the spatio-temporal encoding matrix includes short-term sensitive features and long-term trend features, and the dynamic load information includes historical load information and real-time load information;

[0022] Among them, the enhanced spatio-temporal position encoding is represented by the following formula:

[0023]

[0024] In the formula, t represents the time step, pos represents the position dimension, PE(t, pos) represents the spatio-temporal position encoding, d represents the model dimension, α represents the load dynamic enhancement coefficient, l t represents the real-time load information, represents the historical load mean, represents the historical load standard deviation.

[0025] Furthermore, the step of inputting the spatio-temporal encoding matrix into a preset load prediction model to obtain a load prediction result includes:

[0026] Input the spatio-temporal encoding matrix into a preset load prediction model, where the load prediction model includes a gated recurrent unit, a long short-term memory neural network, and a time-series load prediction module. Among them, the gated recurrent unit and the long short-term memory neural network are in parallel, and are in series with the time-series load prediction module after being in parallel;

[0027] Extract short-term features from the spatio-temporal encoding matrix through the gated recurrent unit to obtain short-term fluctuation features;

[0028] Extract long-term features from the spatio-temporal encoding matrix through the long short-term memory neural network to obtain long-term trend features;

[0029] Perform weighted fusion on the short-term fluctuation features and the long-term trend features to obtain a fused feature matrix;

[0030] Input the fused feature matrix into the time-series load prediction module to obtain a load prediction result, where the load prediction result includes a short-term prediction result and a long-term prediction result.

[0031] Furthermore, the step of dynamically regulating the load of distributed new energy connected to the power grid according to the real-time operation data and the load prediction result includes:

[0032] Perform minute-level response scheduling on the load of distributed new energy connected to the power grid according to the real-time operation data and the fused feature matrix;

[0033] Perform hourly-level optimization scheduling on the load of distributed new energy connected to the power grid according to the fused feature matrix.

[0034] Further, the step of performing minute - level response scheduling on the load of distributed new energy connected to the grid according to the real - time operation data and the fusion feature matrix includes:

[0035] Judging whether an abnormal event occurs in the distributed new energy connected to the grid according to the short - term fluctuation characteristics in the real - time operation data and the fusion feature matrix;

[0036] Executing the corresponding minute - level response scheduling strategy according to the event type of the abnormal event;

[0037] Among them, the event types include sudden drop in new energy output, load mutation, and line overload, and the corresponding minute - level response scheduling strategies for the event types are starting standby power supply, load transfer, and dynamic load shedding respectively.

[0038] Further, the step of performing hourly - level optimal scheduling on the load of distributed new energy connected to the grid according to the fusion feature matrix includes:

[0039] Establishing a multi - objective optimal scheduling model with the minimization of the comprehensive operation cost and the maximization of new energy consumption as the objective functions according to the long - term trend characteristics in the fusion feature matrix;

[0040] Solving the multi - objective optimal scheduling model to obtain the hourly - level optimal scheduling strategy.

[0041] In a second aspect, the present invention provides a load dynamic regulation system for distributed new energy connected to the grid, and the system includes:

[0042] A data pre - processing module, configured to obtain multi - source data of distributed new energy connected to the grid, and perform dynamic alignment and feature extraction on the multi - source data to obtain a multi - source feature tensor, where the multi - source data includes new energy power generation data, meteorological data, and load data;

[0043] A spatio - temporal encoding module, configured to input the multi - source feature tensor into a preset spatio - temporal encoding model to obtain a spatio - temporal encoding matrix, and the spatio - temporal encoding model is constructed based on a deep learning model of the multi - head attention mechanism;

[0044] A load prediction module, configured to input the spatio - temporal encoding matrix into a preset load prediction model to obtain a load prediction result, and the load prediction model is obtained based on the cross - scale joint modeling of the gated recurrent unit and the long short - term memory neural network;

[0045] A dynamic regulation module, configured to obtain the real - time operation data of distributed new energy connected to the grid, and perform dynamic regulation on the load of distributed new energy connected to the grid according to the real - time operation data and the load prediction result.

[0046] The present invention provides a method and system for dynamically regulating the load of distributed new energy connected to the power grid. Through multi-source data alignment optimization, the present invention can solve the problem of inconsistent time granularity, reduce feature loss and noise interference. Through an adaptive spatio-temporal attention template and a spatio-temporal coding model, it can effectively separate short-term fluctuations and long-term trends. By combining a gated recurrent unit and a long short-term memory neural network, the ability of time series modeling is enhanced, the fusion of short-term features and long-term features can be realized, the accuracy of the load prediction result can be improved, and through short-term response scheduling and long-term optimization scheduling, the efficiency of scheduling response can be improved, providing high-timeliness support for the intelligent scheduling of a new energy high-penetration power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic flowchart of a method for dynamically regulating the load of distributed new energy connected to the power grid in an embodiment of the present invention;

[0048] Figure 2 is a schematic structural diagram of a system for dynamically regulating the load of distributed new energy connected to the power grid in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] Please refer to Figure 1 , a method for dynamically regulating the load of distributed new energy connected to the power grid proposed in the first embodiment of the present invention, which includes steps S10 to S40:

[0051] Step S10, obtain multi-source data of distributed new energy connected to the power grid, and perform dynamic alignment and feature extraction on the multi-source data to obtain a multi-source feature tensor, where the multi-source data includes new energy power generation data, meteorological data, and load data;

[0052] Step S20, input the multi-source feature tensor into a preset spatio-temporal coding model to obtain a spatio-temporal coding matrix, where the spatio-temporal coding model is constructed based on a deep learning model of a multi-head attention mechanism;

[0053] Step S30, input the spatio-temporal coding matrix into a preset load prediction model to obtain a load prediction result, where the load prediction model is obtained based on cross-scale joint modeling of a gated recurrent unit and a long short-term memory neural network;

[0054] Step S40, obtaining real-time operation data of the distributed renewable energy access to the power grid, and dynamically regulating the load of the distributed renewable energy access to the power grid according to the real-time operation data and the load forecast result.

[0055] The present invention provides a method for dynamically regulating the load of a power grid for large-scale grid connection of distributed renewable energy sources. When performing grid regulation, it is necessary to analyze various data of the power grid. Since the data sources of the power grid are diverse, such as renewable energy output data, meteorological data, and load data, and there are differences in time scales and sampling frequencies among the multi-source data, traditional methods are difficult to effectively align and fuse. Therefore, this embodiment provides a method for dynamically aligning and extracting features from the multi-source data of the power grid, so as to unify the time granularity of the multi-source data and extract cross-scale features.

[0056] In this embodiment, multi-source data is obtained in units of one day, and the multi-source data includes renewable energy power generation data, meteorological data and load data, among which renewable energy power generation data is minute-level data, including photovoltaic and wind power power series, meteorological data is hour-level data, including temperature, wind speed and irradiance, and load data is data collected at fixed intervals, such as 15-minute intervals, and load data may be missing or abnormal due to holidays or emergencies. First, dynamic segmented statistics are performed on the multi-source data. When performing segmented statistics, segmented statistics are performed on renewable energy power generation data and meteorological data according to the sliding window mechanism, and linear interpolation is used to fill missing values ​​for load data, and 3σ principle and sliding window filtering are used to remove abnormal values. Specifically, a sliding window is used for renewable energy power generation data, for example, the window is 60 minutes and the step size is 15 minutes, the mean of the renewable energy power generation data is calculated, and the average output level within the hourly window is obtained; a sliding window is used for meteorological data, for example, the window is 3 hours and the step size is 1 hour, the trend slope is calculated, and the long-term change trend of meteorological conditions is quantified. The cubic spline interpolation method is used to fill in the missing values, so as to keep the second-order derivative of the load curve continuous and avoid the "sawtooth" effect of linear interpolation. The ADF test is passed to calculate the stability index of the load data.

[0057] Meanwhile, perform timeline alignment on these multi-source data, unify data with different frequencies to the same time granularity. Taking the minute-level time granularity as an example, since the new energy power generation data is at the minute level, it remains unchanged. For meteorological data and load data, perform upsampling. Through linear interpolation, fill in the meteorological data and load data to make the data reach the minute-level time granularity. Of course, the multi-source data can also be aligned at the hour level. At this time, the meteorological data remains unchanged. For the new energy power generation data and load data, perform downsampling, such as performing mean statistics according to the hour window. After dynamically aligning and extracting features from the multi-source data, an aligned multi-source feature tensor is obtained. The multi-source feature tensor is three-dimensional data, and its format is (time step × data source × feature). Here, the features include the mean of the new energy power generation data, the trend slope of the meteorological data, and the stationarity index of the load data.

[0058] After obtaining the multi-source feature tensor, input it into the pre-trained spatio-temporal encoding model to perform cross-scale separation of short-term and long-term features and encode the features. The specific steps include:

[0059] According to the multi-head attention separation mechanism of the spatio-temporal encoding model, perform feature splitting on the multi-source feature tensor;

[0060] Enhance the spatio-temporal position encoding according to the dynamic load information to obtain a spatio-temporal encoding matrix. The spatio-temporal encoding matrix includes short-term sensitive features and long-term trend features. The dynamic load information includes historical load information and real-time load information.

[0061] In this embodiment, the spatio-temporal encoding model is constructed based on the deep learning model Transformer with a multi-head attention mechanism. Among them, the number of short-term attention heads is the same as the number of long-term attention heads, for example, both are 4 heads. Through multi-head attention, perform feature splitting and splicing on the multi-source feature tensor, and output the corresponding position encoding matrix. Feature splitting means splitting the multi-source feature tensor into short-term features and long-term features. The short-term features are related to the new energy power generation data and load data, such as the new energy fluctuation variance, load instantaneous value, etc. The meteorological data and load data are related to the long-term features, such as the meteorological trend slope, load mean, load stationarity index, etc. Specific features can be flexibly selected according to the actual situation during model construction and are not overly limited here. Through the multi-head attention separation mechanism, it is possible to capture both the short-term fluctuations and long-term trends of the grid load simultaneously.

[0062] Since conventional positional encoding only relies on fixed sine functions, for the integration of distributed new energy into the power grid, the changes in its actual load will also affect the fluctuations in new energy output. In order to improve the model's ability to capture the multi-time scale characteristics of the power grid, in this embodiment, traditional positional encoding is fused with dynamic load information to achieve enhanced spatial positional encoding. The enhanced spatio-temporal positional encoding formula is as follows:

[0063]

[0064] In the formula, t represents the time step, that is, the absolute position of the current moment in the sequence; pos represents the position dimension, that is, the dimension index of the positional encoding; d represents the model dimension, that is, the hidden layer dimension of the model; PE(t, pos) represents the spatio-temporal positional encoding, and α represents the load dynamic enhancement coefficient, which is used to adjust the contribution weight of the load information in the positional encoding to avoid the load characteristics dominating the original position information. Preferably, α is set to 0.3; l t represents the real-time load information, that is, the total load value of the power grid at the current moment. represents the historical load mean, which is used for normalization to eliminate the dimension difference and ensure the compatibility of power grid data in different regions. represents the historical load standard deviation, which is used to reflect the load fluctuation intensity.

[0065] In the above formula, the sin function part is the traditional positional encoding, and the latter part is the load dynamic enhancement term. It can be seen that the load dynamic enhancement term is the product of the load dynamic enhancement coefficient and the normalized load deviation. Therefore, the load dynamic enhancement term actually integrates the normalized load deviation into the positional encoding, enabling the model to distinguish between peak and valley load periods and capture abnormal fluctuations. For example, during peak hours, , at this time the encoding value increases, strengthening the model's attention to the high-load state. When , that is, when there is an outlier that does not satisfy the 3σ principle, the encoding value will deviate significantly from the baseline, and at this time, the anomaly detection mechanism can be triggered. In this embodiment, minute-level temporal dependencies (such as new energy output fluctuations) are captured through the sine term, and hourly-level trends are implied through the load enhancement term, that is, the load cycle law is reflected through the mean and standard deviation of historical loads, thereby achieving the multi-time scale fusion of high-frequency and low-frequency features. The spatio-temporal positional encoding enhancement design of this embodiment significantly improves the spatio-temporal modeling ability of the Transformer model in complex power grid scenarios through the organic combination of static position signals and dynamic load characteristics, providing theoretical guarantee and technical support for the precise dispatching of high-penetration new energy power grids.

[0066] In order to eliminate abnormal fluctuations and provide a reliable reference for the attention mechanism of the Transformer model, in a preferred embodiment, before inputting the multi-source feature tensor into the Transformer model, the present invention generates a spatio-temporal attention template by performing hierarchical feature engineering on the multi-source feature tensor, and embeds the spatio-temporal attention template into the Transformer model, thereby improving the attention of the Transformer model to short-term fluctuations in the power grid. The specific steps include:

[0067] Perform velocity hierarchical feature engineering on the multi-source feature tensor to obtain a velocity hierarchical feature matrix;

[0068] Input the velocity hierarchical feature matrix into a preset weight mapping model to obtain a spatio-temporal attention template, and embed the spatio-temporal attention template into a preset spatio-temporal encoding model. The weight mapping model is constructed based on a multi-layer perceptron.

[0069] In this embodiment, first, velocity hierarchical feature engineering is performed on the multi-source feature tensor. For new energy power generation data, according to the minimum value and hourly average of hourly new energy power generation data, calculate its fluctuation variance. According to the trend slope of meteorological data, use the sigmoid function to calculate the trend sensitivity:

[0070]

[0071] In the formula, represents the trend sensitivity at time t, represents the trend slope at time t, represents the slope threshold, k represents the sensitivity scaling factor, and t represents the time step, which can also be understood as the current moment.

[0072] For load data, perform an ADF test to calculate its stationarity index. If this index exists in the multi-source feature tensor, it can also be directly extracted and used. Finally, combine the fluctuation variance, trend sensitivity, and stationarity index to obtain a velocity hierarchical feature matrix.

[0073] Then input the velocity hierarchical feature matrix into a preset weight mapping model to obtain a spatio-temporal attention template. The specific steps include:

[0074] Map the velocity hierarchical feature matrix into attention weights through the weight mapping model. The attention weights include short-term attention weights and long-term attention weights;

[0075] Smoothly optimize the attention weights through a dynamic programming algorithm to generate a spatio-temporal attention template.

[0076] In this embodiment, the weight mapping model is constructed based on a multi-layer perceptron (MLP), trained with historical data, and maps the input velocity stratification feature matrix to output the short-term attention weight sequence in the Transformer model. Here, the sum of the short-term attention weights and the long-term attention weights is 1. Therefore, the long-term attention weight sequence can be directly obtained from the short-term attention weight sequence.

[0077] To improve the stability of the weight sequence and provide the reliability for subsequent decision-making analysis, in this embodiment, the weight sequence is smoothed and optimized using a dynamic programming algorithm, with the objective function being the minimization of the weight difference between adjacent time steps and the deviation between the smoothed weight and the original estimated weight:

[0078]

[0079] In the formula, represents the short-term attention weight smoothed at time step t, represents the original estimated weight output by the model, T represents the time window length, which is T = 60 if the step size is 1 hour, and λ represents the smoothing coefficient.

[0080] The above formula includes a smoothing term and a fidelity term. Among them, the smoothing term is: , and the fidelity term is: .

[0081] The purpose of the smoothing term is to minimize the weight difference between adjacent time steps, avoid drastic jumps, thereby suppressing noise interference and enhancing the model's robustness to short-term fluctuations. The role of the fidelity term is to limit the deviation between the smoothed weight and the original estimated weight, preventing over-smoothing from causing the model to lose key fluctuation information. The smoothing coefficient serves as the balancing weight for controlling the smoothing term and the fidelity term. Set the initial weight value as the original estimated weight output by the MLP model for the first time , to avoid initial jumps, and then perform dynamic programming iterative solution on the objective function to obtain the smoothed short-term attention weight sequence. Based on the constraint relationship between the short-term attention weights and the long-term attention weights, obtain the long-term attention weight sequence, and finally output the spatio-temporal attention template containing the short-term attention weight sequence and the long-term attention weight sequence.

[0082] In this embodiment, the velocity stratification feature engineering is used to quantify the change velocities at different time scales, thereby adjusting the attention weights in the new energy scenario. Embedding the generated spatio-temporal attention template into the Transformer model can effectively improve the attention focusing accuracy of the Transformer model for critical periods (such as sudden drops in new energy output), thus providing more accurate data support for subsequent load forecasting.

[0083] In the above embodiments, the spatio-temporal encoding matrix output by the Transformer model actually contains short-term features and long-term features. Assuming that the spatio-temporal encoding matrix is 256-dimensional, the first 128 dimensions are short-term features, and the last 128 dimensions are long-term features. In order to overcome the problem of low prediction accuracy of a single model caused by the coupling effect of data at different time scales, in a preferred embodiment, the present invention constructs a load prediction model through a cross-scale joint modeling method of a gated recurrent unit (GRU) and a long short-term memory neural network (LSTM) to achieve collaborative processing of short-term features and long-term features.

[0084] In this embodiment, the GRU and LSTM in the load prediction model are in parallel. When the spatio-temporal encoding matrix is input into the trained load prediction model, the spatio-temporal encoding matrix will be separately input into the GRU and LSTM. The GRU and LSTM extract short-term features and long-term features respectively according to the matrix dimensions. In the GRU, the retention ratio of information is controlled by an update gate, and the historical memory is reset by a reset gate to achieve short-term feature processing. Its output is short-term fluctuation features at the minute level, such as new energy output variance, load instantaneous change rate, meteorological difference volatility, etc.; in the LSTM, long-term feature processing is achieved through cell state update, and its output is long-term trend features at the hour level, such as meteorological trend slope, daily load cycle pattern, long-term prediction deviation of new energy output, etc.

[0085] For the short-term fluctuation features output by the GRU and the long-term trend features output by the LSTM, they are weighted and fused through a cross-scale fusion gate to obtain a fused feature matrix. When performing feature fusion, the weight value can be a preset value or determined by dynamic weight calculation. When using dynamic weights, the sigmoid function is used to calculate the dynamic weights:

[0086]

[0087] In the formula, represents the weight of the short-term fluctuation features, represents the preset dynamic weight matrix, GRU t represents the short-term fluctuation features, LSTM t represents the long-term trend features.

[0088] Then, through dynamic weighted fusion, the fused feature matrix is obtained:

[0089]

[0090] In the formula, Output t represents the fused feature matrix, and its format is (time step × feature dimension).

[0091] Then, the fused feature matrix is input into the time-series load forecasting module for load forecasting. The time-series load forecasting module can be a fully connected layer or a lightweight time-series forecasting module constructed based on linear regression or a shallow neural network, so as to generate the final load forecasting result.

[0092] In a preferred embodiment, the time-series load forecasting module includes a short-term forecasting sub-module and a long-term forecasting sub-module. Among them, the short-term forecasting sub-module is constructed based on a lightweight time-series convolutional network (TCN), and the long-term forecasting sub-module is constructed based on autoregressive integrated moving average (ARIMA). When making a forecast, the fused feature matrix is input into the two sub-modules respectively. The sub-modules will extract short-term fluctuation features and long-term trend features through different dimensions of the matrix, so as to realize load forecasting at different time scales. The cross-scale joint modeling method of this embodiment overcomes the problem of insufficient capture ability of a single model for minute-level mutations and hour-level trends, and improves the accuracy of model prediction.

[0093] According to the load forecasting result and combined with the real-time operation data of distributed new energy connected to the power grid, the load of distributed new energy connected to the power grid can be dynamically regulated. When regulating, conventional power grid regulation methods can be referred to, and the power grid can be regulated according to the predicted load and the current operation data, so that the power grid operation strategy matches the load forecasting result.

[0094] In order to improve the response time of power grid regulation and achieve cross-scale dynamic scheduling decisions, in a preferred embodiment, the present invention provides a cross-scale dynamic regulation method, and the specific steps include:

[0095] Perform minute-level response scheduling on the load of distributed new energy connected to the power grid according to the real-time operation data and the fused feature matrix;

[0096] Perform hour-level optimization scheduling on the load of distributed new energy connected to the power grid according to the fused feature matrix.

[0097] In this embodiment, the dynamic regulation of the power grid load includes minute-level emergency response and hour-level optimization scheduling. Among them, the steps of the minute-level emergency response include:

[0098] Judge whether an abnormal event occurs in the distributed new energy connected to the power grid according to the real-time operation data and the short-term fluctuation features in the fused feature matrix;

[0099] Execute the corresponding minute-level response scheduling strategy according to the event type of the abnormal event;

[0100] Among them, the event types include sudden drop in new energy output, load mutation, and line overload, and the corresponding minute-level response scheduling strategies for the event types are starting standby power supply, load transfer, and dynamic load shedding respectively.

[0101] In this embodiment, the data used for the dispatching decision of minute-level emergency response includes the short-term fluctuation characteristics in the fusion feature matrix and the real-time operation data of the power grid. The real-time operation data further includes real-time load data, the state of charge (SOC) of the energy storage device, and the real-time line load rate. Then, based on the short-term fluctuation characteristics and the real-time operation data of the power grid, it is determined whether an abnormal event occurs in the power grid, and the event types include a sudden drop in new energy output, a load mutation, and a line overload.

[0102] Specifically, according to the variance of new energy output in the short-term fluctuation characteristics, it is determined whether a sudden drop in new energy output event occurs. If it exists and the short-term fluctuation exceeds the long-term benchmark, the standby power supply is activated, that is, the energy storage discharge is activated according to the state of charge (SOC) of the energy storage device. According to the real-time load data and the load instantaneous change rate in the short-term fluctuation characteristics, it is determined whether a load mutation event occurs. If it occurs, load transfer dispatching is performed on the power grid. According to the real-time line load rate, it is determined whether a line overload event occurs. If it occurs, dynamic load shedding dispatching is performed on the power grid. It should be noted that abnormal event determination can also be performed through other short-term fluctuation characteristics and real-time operation data. The specific abnormal event types and corresponding determination methods can be flexibly set according to the actual situation. This embodiment is only a preferred method rather than a specific limitation.

[0103] In a preferred embodiment, the steps of hourly-level optimal dispatching include:

[0104] Based on the long-term trend characteristics in the fusion feature matrix, a multi-objective optimal dispatching model is established with the minimization of the comprehensive operation cost and the maximization of new energy consumption as the objective function.

[0105] The multi-objective optimal dispatching model is solved to obtain the hourly-level optimal dispatching strategy.

[0106] In this embodiment, the multi-objective optimal dispatching model is established based on the long-term trend characteristics to achieve hourly-level optimal dispatching. Among them, the multi-objective optimal dispatching model takes the minimization of the comprehensive operation cost and the maximization of new energy consumption as the objective function, and its formula is expressed as:

[0107]

[0108] In the formula, represents the long-term trend characteristics at the t time step, represents the meteorological prediction data at the t time step, and this meteorological prediction value can be obtained through the meteorological prediction department. represents the historical load data at the t time step, It represents the dispatching instruction vector, which is the power distribution plan of each power source (such as coal-fired units, new energy power stations, energy storage systems, etc.) to be determined in the optimization model within the specified time period. θ represents the weight of the meteorological matching term, which is used to emphasize the dominant influence of meteorological prediction on dispatching and prevent insufficient output caused by meteorological mutations. γ represents the weight of the historical fitting term, which is used to balance the innovation of the model and historical experience and avoid overfitting new data. η represents the sparsity constraint coefficient, which is used to control the sparsity of the dispatching instruction. The larger the value, the fewer the starting and stopping units. T represents the time window length, which is used to cover the typical cycle of power grid dispatching, and θ + γ = 1. Preferably, θ is set to 0.8 and γ is set to 0.2. It should be noted that in multiple formulas of the present invention, there are the time window length T and the time step t. The specific values are different in different formulas, but their meanings are the same, and they will not be distinguished by different parameters here.

[0109] In the above objective function, it includes three parts, namely the meteorological matching term, the historical fitting term, and the sparsity constraint. Among them, the meteorological matching term is the square of the L2 norm between the long-term trend feature and the meteorological prediction data:

[0110]

[0111] The objective function minimizes the difference between the long-term trend features (such as the daily load cycle, new energy output deviation) and the meteorological prediction data (such as temperature trend, wind speed change), in order to ensure that the dispatching plan conforms to the changes in meteorological conditions, prevent insufficient output caused by meteorological mutations, and reduce the unplanned costs caused by weather mutations. At the same time, using the square of the L2 norm in the meteorological matching term can ensure that the objective function is a convex function, so there is a global optimal solution, and higher weights are given to large deviations (because the error will grow quadratically), prompting the model to pay more attention to the overall matching of meteorological trends.

[0112] The historical fitting term is the L1 norm between the long-term trend feature and the historical load data:

[0113]

[0114] The historical fitting term is to constrain the deviation of the dispatching plan from the historical experience mode (such as the load peak during holidays), and avoid the over-investment cost caused by over-reliance on prediction.

[0115] The sparsity constraint is the L1 norm of the dispatching instruction vector:

[0116]

[0117] Through the sparsity constraint, the frequent starting and stopping of units can be reduced, thereby reducing the loss cost of equipment.

[0118] Furthermore, since the long-term trend feature includes the long-term prediction deviation of new energy output (such as the daily cumulative deviation of wind power). During the optimization process, the model can dynamically adjust the proportion of thermal power and new energy output by minimizing this deviation. For example, if the predicted wind power output is higher than the actual value (the deviation is negative), the thermal power reserve is increased; if the prediction is low (the deviation is positive), the thermal power is reduced, and the new energy is preferentially absorbed. Although the curtailment term is not explicitly included in the objective function, the L2 norm constraint of the meteorological matching term can indirectly reduce the curtailment rate; in addition, the sparsity constraint prompts the energy storage to charge during the peak of new energy output (low electricity price) and discharge during the trough (reducing the thermal power demand), indirectly improving the utilization rate of new energy. Therefore, by iteratively solving the objective function, an hourly optimal dispatching strategy that minimizes the comprehensive operation cost and maximizes the new energy consumption can be obtained.

[0119] A load dynamic regulation method for distributed new energy connected to the power grid provided by this embodiment. Through multi-source data alignment optimization, the present invention solves the problem of inconsistent time granularity, reduces feature loss and noise interference. Through the adaptive spatio-temporal attention template and spatio-temporal coding model, the short-term fluctuations and long-term trends are effectively separated. By combining the gated recurrent unit and the long short-term memory neural network, the temporal modeling ability is enhanced, the fusion of short-term and long-term features is realized, the accuracy of the load prediction result is improved, and through the short-term response scheduling and long-term optimal scheduling, the efficiency of the scheduling response is improved, providing high-timeliness support for the intelligent scheduling of the new energy high-penetration power grid.

[0120] Please refer to Figure 2 , based on the same inventive concept, a load dynamic regulation system for distributed new energy connected to the power grid proposed in the second embodiment of the present invention includes:

[0121] A data preprocessing module 10, configured to obtain multi-source data of distributed new energy connected to the power grid, and perform dynamic alignment and feature extraction on the multi-source data to obtain a multi-source feature tensor, where the multi-source data includes new energy generation data, meteorological data, and load data;

[0122] A spatio-temporal coding module 20, configured to input the multi-source feature tensor into a preset spatio-temporal coding model to obtain a spatio-temporal coding matrix, where the spatio-temporal coding model is constructed based on a deep learning model of the multi-head attention mechanism;

[0123] A load prediction module 30, configured to input the spatio-temporal coding matrix into a preset load prediction model to obtain a load prediction result, where the load prediction model is obtained based on the cross-scale joint modeling of the gated recurrent unit and the long short-term memory neural network;

[0124] A dynamic regulation module, which is used to obtain the real-time operation data of distributed new energy connected to the power grid, and dynamically regulate the load of distributed new energy connected to the power grid according to the real-time operation data and the load prediction result.

[0125] The technical features and technical effects of the load dynamic regulation system for distributed new energy connected to the power grid proposed in the embodiments of the present invention are the same as those of the method proposed in the embodiments of the present invention, and will not be elaborated here. Each module in the above-mentioned load dynamic regulation system for distributed new energy connected to the power grid can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in the processor in the computer device in the form of hardware or be independent of it, or can be stored in the memory in the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the above-mentioned modules.

[0126] In summary, the embodiments of the present invention propose a load dynamic regulation method and system for distributed new energy connected to the power grid. The method obtains multi-source data of distributed new energy connected to the power grid, and performs dynamic alignment and feature extraction on the multi-source data to obtain a multi-source feature tensor. The multi-source data includes new energy power generation data, meteorological data and load data; the multi-source feature tensor is input into a preset spatio-temporal coding model to obtain a spatio-temporal coding matrix, and the spatio-temporal coding model is constructed based on a deep learning model with a multi-head attention mechanism; the spatio-temporal coding matrix is input into a preset load prediction model to obtain a load prediction result, and the load prediction model is obtained based on the cross-scale joint modeling of a gated recurrent unit and a long short-term memory neural network; the real-time operation data of distributed new energy connected to the power grid is obtained, and the load of distributed new energy connected to the power grid is dynamically regulated according to the real-time operation data and the load prediction result. The present invention solves the problem of inconsistent time granularity through multi-source data alignment optimization, reduces feature loss and noise interference, effectively separates short-term fluctuations and long-term trends through an adaptive spatio-temporal attention template and a spatio-temporal coding model, enhances the time series modeling ability through the combination of a gated recurrent unit and a long short-term memory neural network, realizes the fusion of short-term features and long-term features, improves the accuracy of the load prediction result, and improves the efficiency of the scheduling response through short-term response scheduling and long-term optimization scheduling, providing high-timeliness support for the intelligent scheduling of a new energy high-penetration power grid.

[0127] The various embodiments in this specification are described in a progressive manner. For the parts that are the same or similar in each embodiment, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiment. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0128] The above embodiments only represent several preferred embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claimed rights.

Claims

1. A method for dynamically controlling the load of distributed renewable energy connected to a power grid, characterized in that: include: Acquire multi-source data of distributed renewable energy access to the power grid, and dynamically align and extract features from the multi-source data to obtain a multi-source feature tensor, wherein the multi-source data includes renewable energy power generation data, meteorological data, and load data; Inputting the multi-source feature tensor into a preset spatiotemporal coding model to obtain a spatiotemporal coding matrix, wherein the spatiotemporal coding model is constructed based on a deep learning model of a multi-head attention mechanism; Inputting the spatiotemporal coding matrix into a preset load forecasting model to obtain a load forecasting result, wherein the load forecasting model is obtained based on cross-scale joint modeling of a gated recurrent unit and a long short-term memory neural network; Acquire real-time operation data of distributed renewable energy access to the power grid, and dynamically regulate the load of distributed renewable energy access to the power grid according to the real-time operation data and the load forecast result; The step of inputting the multi-source feature tensor into a preset spatiotemporal coding model to obtain a spatiotemporal coding matrix includes: According to the multi-head attention separation mechanism of the spatiotemporal coding model, feature decomposition is performed on the multi-source feature tensor; Performing spatiotemporal position coding enhancement according to the dynamic load information to obtain a spatiotemporal coding matrix, wherein the spatiotemporal coding matrix includes short-term sensitive features and long-term trend features, and the dynamic load information includes historical load information and real-time load information; Among them, the enhanced spatiotemporal position coding is expressed by the following formula: In the formula, t represents the time step, pos represents the position dimension, and PE (t,pos) represents the spatiotemporal position encoding, d represents the model dimension, α represents the load dynamic enhancement coefficient, l t Indicates real-time load information, represents the historical load mean, Represents the historical load standard deviation.

2. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 1, characterized in that: Before the step of inputting the multi-source feature tensor into a preset spatiotemporal coding model to obtain a spatiotemporal coding matrix, the method further includes: Performing speed-layered feature engineering on the multi-source feature tensor to obtain a speed-layered feature matrix; The speed hierarchical feature matrix is ​​input into a preset weight mapping model to obtain a spatiotemporal attention template, and the spatiotemporal attention template is embedded into a preset spatiotemporal coding model. The weight mapping model is constructed based on a multi-layer perceptron.

3. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 2, characterized in that: The step of performing speed hierarchical feature engineering on the multi-source feature tensor to obtain a speed hierarchical feature matrix comprises: respectively calculating the fluctuation variance of the new energy power generation data, the trend sensitivity of the meteorological data and the stability index of the load data; The fluctuation variance, the trend sensitivity and the stability index are combined to obtain a speed stratified feature matrix.

4. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 2, characterized in that: The step of inputting the speed hierarchical feature matrix into a preset weight mapping model to obtain a spatiotemporal attention template comprises: Mapping the speed hierarchical feature matrix into attention weights through a weight mapping model, wherein the attention weights include short-term attention weights and long-term attention weights; The attention weights are smoothly optimized through a dynamic programming algorithm to generate a spatiotemporal attention template.

5. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 1, characterized in that: The step of inputting the space-time coding matrix into a preset load forecasting model to obtain a load forecasting result comprises: Inputting the spatiotemporal coding matrix into a preset load forecasting model, wherein the load forecasting model includes a gated recurrent unit, a long short-term memory neural network, and a time series load forecasting module, wherein the gated recurrent unit and the long short-term memory neural network are connected in parallel and then connected in series with the time series load forecasting module after being connected in parallel; Performing short-term feature extraction on the spatiotemporal coding matrix through a gated recurrent unit to obtain short-term fluctuation features; Extracting long-term features from the spatiotemporal coding matrix through a long short-term memory neural network to obtain long-term trend features; Performing weighted fusion on the short-term fluctuation feature and the long-term trend feature to obtain a fusion feature matrix; The fused feature matrix is ​​input into a time series load forecasting module to obtain a load forecasting result, which includes a short-term forecasting result and a long-term forecasting result.

6. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 5, characterized in that: The step of dynamically regulating the load of the distributed renewable energy source connected to the power grid according to the real-time operation data and the load forecast result comprises: According to the real-time operation data and the fusion feature matrix, minute-level response scheduling is performed on the load of the distributed renewable energy access to the power grid; According to the fusion characteristic matrix, the load of the distributed renewable energy connected to the power grid is optimally scheduled at the hourly level.

7. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 6, characterized in that: The step of performing minute-level response scheduling on the load of the distributed renewable energy access to the power grid according to the real-time operation data and the fusion feature matrix comprises: Judging whether an abnormal event occurs when the distributed renewable energy is connected to the power grid based on the real-time operation data and the short-term fluctuation characteristics in the fusion feature matrix; According to the event type of the abnormal event, execute the corresponding minute-level response scheduling strategy; Among them, the event types include sudden drop in new energy output, sudden load change and line overload, and the minute-level response scheduling strategies corresponding to the event types are starting backup power supply, load transfer and dynamic load shedding.

8. The method for dynamic load control of distributed renewable energy access to a power grid according to claim 6, characterized in that: The step of performing hourly optimization scheduling on the load of the distributed renewable energy connected to the power grid according to the fusion characteristic matrix comprises: According to the long-term trend characteristics in the fusion feature matrix, a multi-objective optimization scheduling model is established with minimization of comprehensive operating costs and maximization of new energy consumption as objective functions; The multi-objective optimization scheduling model is solved to obtain an hourly optimization scheduling strategy.

9. A load dynamic control system for accessing a distributed renewable energy grid, characterized in that: include: A data preprocessing module is used to obtain multi-source data of distributed renewable energy access to the power grid, and dynamically align and extract features from the multi-source data to obtain a multi-source feature tensor, wherein the multi-source data includes renewable energy power generation data, meteorological data, and load data; A spatiotemporal coding module is used to input the multi-source feature tensor into a preset spatiotemporal coding model to obtain a spatiotemporal coding matrix, wherein the spatiotemporal coding model is constructed based on a deep learning model of a multi-head attention mechanism; comprising: According to the multi-head attention separation mechanism of the spatiotemporal coding model, feature decomposition is performed on the multi-source feature tensor; Performing spatiotemporal position coding enhancement according to the dynamic load information to obtain a spatiotemporal coding matrix, wherein the spatiotemporal coding matrix includes short-term sensitive features and long-term trend features, and the dynamic load information includes historical load information and real-time load information; Among them, the enhanced spatiotemporal position coding is expressed by the following formula: In the formula, t represents the time step, pos represents the position dimension, and PE (t,pos) represents the spatiotemporal position encoding, d represents the model dimension, α represents the load dynamic enhancement coefficient, l t Indicates real-time load information, represents the historical load mean, represents the standard deviation of historical load; A load forecasting module, used for inputting the spatiotemporal coding matrix into a preset load forecasting model to obtain a load forecasting result, wherein the load forecasting model is obtained based on cross-scale joint modeling of a gated recurrent unit and a long short-term memory neural network; The dynamic control module is used to obtain real-time operation data of distributed renewable energy access to the power grid, and dynamically control the load of distributed renewable energy access to the power grid based on the real-time operation data and the load forecast results.

Images