Power grid load prediction method based on multi-source data and hybrid neural network

Abstract

The invention provides a power grid load prediction method based on multi-source data and a hybrid neural network, and belongs to the technical field of load prediction. The method includes: acquiring historical data of load and associated factors of a target power grid; re-sampling the historical data according to a preset sampling frequency to obtain re-sampled data; noise reduction processing is carried out on high-frequency fluctuation type data in the resampling data, and smoothing processing is carried out on step change type data in the sampling data; adding time features and weather forecast data to each group of data in the resampling data, and performing normalization processing; a load prediction model is constructed, and training of the load prediction model is completed by using the resampling data; and performing power grid load prediction by using the load prediction model. According to the invention, by accurately predicting the load demand, reasonable scheduling and balance of the load can be realized, the condition of excessive or insufficient load is avoided, and comprehensive monitoring and management of energy consumption can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of load forecasting, and particularly relates to a power grid load forecasting method based on multi-source data and a hybrid neural network. Background Art

[0002] With the rapid economic development leading to an expanding demand for energy electricity, the power generation balance fluctuation increases, bringing greater requirements for the transmission capacity and peak regulation capacity of the power grid network. In the case of a large number of distributed energy sources (such as rooftop photovoltaic and small-scale wind power) being connected, the power supply and demand characteristics are significantly changed. On the demand side, the in-situ consumption of distributed energy sources causes the power grid load to fluctuate frequently. For example, when the output of distributed photovoltaics increases, the power consumption of users from the power grid decreases, resulting in a reduction in the power grid load value.

[0003] The current load forecasting is basically based on the fixed interruptible load amount fitted from historical data, and does not fully consider the real-time load elasticity under the penetration of new energy. Especially in extreme weather or scenarios with sudden changes in new energy output, the existing methods will have problems with load forecasting deviation. Therefore, it is urgent to construct a short-term load forecasting model considering the influence of distributed energy sources to support dynamic and accurate load scheduling and resource optimization decisions. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a power grid load forecasting method based on multi-source data and a hybrid neural network.

[0005] To solve one or part or all of the above technical problems, the technical solutions adopted by the present invention are as follows: A power grid load forecasting method based on multi-source data and a hybrid neural network, comprising: obtaining historical data of the load and associated factors of the target power grid; resampling the historical data according to a preset sampling frequency to obtain resampled data; performing noise reduction processing on the high-frequency fluctuation type data in the resampled data, and performing smoothing processing on the step change type data in the sampled data; adding time features and weather forecast data to each group of data in the resampled data, and performing normalization processing; constructing a load forecasting model, and training the load forecasting model using the resampled data; and performing power grid load forecasting using the load forecasting model.

[0006] Further, the load forecasting model includes multiple forecasting modules and an output module connected to the output ends of all the forecasting modules; each forecasting module includes an input layer, a CNN module, an LSTM module, and an attention mechanism layer connected in sequence; the input layer is used to input the high-frequency fluctuation data; the LSTM module is used to process the concatenated features of the step-change data, the time features, the weather forecast data, and the output features of the CNN module; the output module is used to perform weighted fusion on the output features of all the forecasting modules and output the hourly load forecasting results.

[0007] Further, different CNN modules, different LSTM modules, and different attention mechanism layers use different parameters.

[0008] Further, at least one LSTM module is provided with LSTM layers that are all forward, at least one LSTM module is provided with LSTM layers that are all backward, and at least one LSTM module is provided with both forward and backward LSTM layers.

[0009] Further, the output module includes a first fully connected layer, a second fully connected layer, and a third fully connected layer; the first fully connected layer is used to dynamically adjust the weights of different forecasting modules according to the prediction errors of the forecasting modules; the second connected layer is used to perform weighted fusion on the output features of the attention mechanism layer according to the weights of different forecasting modules; the third fully connected layer is used to map the fused features of the second connected layer to the load forecasting results.

[0010] Further, the attention mechanism layer is used to calculate the attention weights for each time-step feature output by the LSTM module and dynamically strengthen the feature expression of the key time steps.

[0011] Further, the method for resampling the historical data includes: if the data at the resampling time point in the historical data is missing or abnormal, interpolation is used for sampling; among them, the cubic spline interpolation method is used for sampling the high-frequency fluctuation data, and the forward filling method is used for sampling the step-change data.

[0012] Further, the autoregressive wavelet denoising method is used to denoise the high-frequency fluctuation data in the resampled data, and the moving average method is used to smooth the step-change data in the sampled data.

[0013] Further, the time features include the annual week number, annual day number, intra-day sampling number, holiday flag, and rest day flag corresponding to the sampling time.

[0014] Further, the method for predicting the power grid load using the load prediction model includes: obtaining the hourly weather forecast data for the next 24 hours and resampling it by means of cubic spline interpolation; inputting the sampled data for the previous 48 hours and the resampled data of the weather forecast data for the next 24 hours into the load prediction model to obtain the hourly load prediction values for the next 24 hours.

[0015] Further, the high-frequency fluctuation type data in the historical data includes load data and weather data, and the step change type data in the historical data includes distributed energy installation capacity data.

[0016] Further, the weather data includes temperature, humidity, wind direction, wind speed, visibility distance, air pressure, solar irradiance, snow depth, precipitation, cloud type.

[0017] Further, the distributed energy installation capacity data includes distributed energy installation capacity data and distributed wind power installation capacity data.

[0018] Further, the sampling interval of the resampled data is 15 minutes.

[0019] Further, the resampled data is normalized by means of Z-score standardization.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: By integrating load data, weather data and distributed energy related data, the present invention can comprehensively cover various dynamic factors affecting the power grid load, especially the relevant parameters affecting the new energy output such as weather type and wind speed, and can accurately capture the consumption impact of new energy power generation on the power grid load.

[0021] By adding multi-dimensional time features, the present invention strengthens the model's ability to identify periodic patterns and emergencies. The annual week number feature helps the model distinguish the load change patterns in different seasons, the intra-day sampling order coding can locate the peak and valley periods of electricity consumption, and the holiday and rest day identifiers can effectively distinguish the load feature differences between working days and special dates.

[0022] The load prediction model provided by the present invention adopts a CNN-LSTM hybrid structure, which can achieve collaborative learning of local features and global time series. The load prediction model includes multiple prediction modules and an output module connected to the output ends of all prediction modules. Each prediction module includes an input layer, a CNN module, an LSTM module, and an attention mechanism layer connected in sequence. Different CNN modules adopt different parameters. By adjusting parameters such as the width and number of convolutional kernels, different CNN modules can extract local features on different time scales, enhancing the sensitivity to intraday load changes and meteorological mutations. Different LSTM modules adopt different parameters. At the same time, different single-direction LSTM layers and bi-directional LSTM layers are used. The single-direction LSTM layer focuses on future trend prediction, and the bi-directional LSTM layer can capture periodic features of long time series such as seasonal changes and holidays, forming a complement. Different attention mechanisms are adopted between different LSTM modules. Through the combination of different attention mechanisms, the dynamic strengthening ability of features in different key time periods can be enhanced. The output module includes multiple fully connected layers. The first fully connected layer dynamically adjusts the weights by analyzing the validation set errors of different prediction modules, which can effectively balance the prediction deviations between prediction modules and avoid the dominant result of a poorly performing model. The second fully connected layer performs weighted fusion on the attention features, retaining the attention selection of each sub-model for time steps and avoiding information loss caused by premature fusion. The third fully connected layer is responsible for the final prediction mapping.

[0023] Through the analysis of historical data and model prediction, the present invention grasps the peak and valley periods of load, energy waste, etc., facilitating the formulation of reasonable load adjustment strategies based on the analysis of prediction data. By accurately predicting the load demand, reasonable scheduling and balancing of the load can be achieved, avoiding the situation of overloaded or underloaded load, and comprehensive monitoring and management of energy consumption can also be realized. Using the prediction results of the present invention, power generation plans and scheduling arrangements can be made to ensure the stability and efficiency of power supply. Brief Description of the Drawings

[0024] The following further elaborates the present invention in detail with reference to the drawings.

[0025] Figure 1 : Schematic diagram of the load prediction model of the present invention. Detailed Embodiments

[0026] To better understand the present invention, the content of the present invention will be further clearly elaborated below in combination with embodiments and drawings. However, the protected content of the present invention is not limited to the following embodiments. In the following description, a large number of specific details are given to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details.

[0027] Embodiment 1: The purpose of this embodiment is to provide a power grid load forecasting method for multi-source data and hybrid neural networks, including: S1. Obtain the historical data of the load and associated factors of the target power grid.

[0028] The historical data of the associated factors includes meteorological data and distributed energy installation capacity data in the region. The meteorological data includes temperature, humidity, wind direction, wind speed, visibility distance, air pressure, solar irradiance, snow depth, precipitation, cloud type; the distributed energy installation capacity data includes distributed energy installation capacity data and distributed wind power installation capacity data.

[0029] The obtained historical data can be divided into high-frequency fluctuation type data and step change type data according to the data change law. The high-frequency fluctuation type data includes load data and meteorological data, and the step change type data includes distributed energy installation capacity data.

[0030] S2. Resample the historical data according to a preset sampling frequency to obtain resampled data.

[0031] If the data at the resampling time point in the historical data is missing or abnormal, interpolation method is used for sampling. For high-frequency fluctuation type data such as load data and meteorological data, cubic spline interpolation method is used for sampling, and for distributed energy installation capacity data, forward filling method is used for sampling. Preferably, the sampling interval for resampling is 15 minutes.

[0032] The distributed energy installation capacity is affected by centralized grid connection, policy drive, grid connection of large-scale new energy projects, etc. The distributed energy installation capacity data shows a stepwise growth characteristic and is not a linear change. Therefore, when the sampled value of the distributed energy installation capacity data is missing, the data at the previous moment in the historical data is directly inherited.

[0033] S3. Denoise the high-frequency fluctuation type data in the resampled data and smooth the step change type data.

[0034] After resampling the high-frequency fluctuation type data by interpolation, new high-frequency noise may be introduced. Therefore, denoising the high-frequency fluctuation type data can eliminate noise and interpolation errors and improve the prediction accuracy. Preferably, the autoregressive wavelet denoising method is used to process the high-frequency fluctuation type data to remove high-frequency noise while retaining key fluctuation characteristics.

[0035] There are mutation data in the step change type data, and the mutation of the data will affect the prediction accuracy of the model. Therefore, the moving average method is used to mildly smooth the step change type data. The window length of the moving average method is 12 hours - 24 hours, which can smooth the mutation data and retain the step mutation characteristics.

[0036] S4. Add time features and meteorological forecast data to each group of data in the resampled data, and perform normalization processing.

[0037] The time features are determined according to the sampling moment of the group of data, including periodic features and event features. The periodic features include the annual week number (1 - 52), the annual day number (1 - 365), and the intra-day sampling ordinal number (0 - 96, with a sampling frequency of 10 / 9 mHz). The event features include the holiday flag (0 or 1) and the rest day flag (0 or 1).

[0038] The meteorological forecast data are the forecast values of the meteorological data at the corresponding moment of the next day, and the specific type is the same as that of the meteorological data in the resampled data. The meteorological forecast data can be obtained by statistically processing the meteorological data of the next day at the current moment in the resampled data.

[0039] Subsequently, the resampled data is normalized using the Z-score normalization method.

[0040] S5. Construct a load forecasting model and use the resampled data to complete the training of the load forecasting model.

[0041] The load forecasting model adopts a CNN-LSTM hybrid neural network, using CNN to process local features and LSTM to capture long-term time series dependencies. Specifically, as Figure 1 shown, the load forecasting model includes multiple prediction modules and an output module connected to the output ends of all prediction modules. Each prediction module includes an input layer, a CNN module, an LSTM module, and an attention mechanism layer connected in sequence. The output module includes multiple fully connected layers.

[0042] The input layer is used to receive high-frequency fluctuating data (including load data and meteorological data).

[0043] The CNN module first uses a 1D convolutional layer to extract local time series features (such as intra-day load fluctuations and short-term meteorological mutations), and sets multiple convolutional kernels in the convolutional layer; then compresses the time series dimension through a max-pooling layer to retain key local features. Different CNN modules use different parameters, such as different convolutional scales, different numbers of convolutional kernels, and / or different convolutional kernel widths.

[0044] The LSTM module directly concatenates the input step-change data (including distributed energy installation data), time features, weather forecast data with the output of the CNN module and inputs them into the LSTM layer to capture long-term dependencies (such as periodic changes, holiday impacts). Different LSTM modules adopt different parameters, such as LSTM layers with different numbers of layers and different directions, and different LSTM cells. At least one LSTM module is provided with a number of LSTM layers all in the forward direction, at least one LSTM module is provided with a number of LSTM layers all in the reverse direction, and at least one LSTM module is simultaneously provided with a number of forward LSTM layers and a number of reverse LSTM layers, enabling the LSTM module to utilize forward and reverse time series information simultaneously and enhancing the ability to capture time series features.

[0045] The attention mechanism layer performs weighted fusion on the output features of the LSTM module, calculates the attention weights for the features of each time step output by the LSTM module, and dynamically strengthens the feature expressions of key time steps (such as meteorological features during high-temperature periods, load features during bad weather). Different attention mechanisms (standard Softmax attention, gated attention, multi-head attention, etc.) are adopted between different LSTM modules.

[0046] The output module performs weighted fusion on the output features of all attention mechanism layers and maps them to the hourly load prediction results for the next 24 hours. The output module includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first fully connected layer is used to dynamically adjust the weights of different prediction modules according to the prediction errors of the prediction modules; the second connection layer is used to perform weighted fusion on the output features of the attention mechanism layer according to the weights of different prediction modules; the third fully connected layer is used to map the fused features of the second connection layer to the load prediction results.

[0047] In this step, after constructing the load prediction model, the resampled data is split into a training set (such as 70%), a validation set (such as 15%), and a test set (such as 15%), and the load prediction model is trained, validated, and tested.

[0048] During the validation process using the validation set, the first fully connected layer analyzes the mean absolute error of the prediction results of each prediction module and dynamically adjusts the weights of different prediction modules according to the prediction errors.

[0049] Preferably, the number of prediction modules is 3 - 5.

[0050] S6. Use the load prediction model to perform power grid load prediction.

[0051] First, obtain the hourly meteorological forecast data for the next 24 hours, and resample it using cubic spline interpolation with the sampling frequency being the same as that of the resampled data. Subsequently, input the sampled data for the previous 48 hours and the resampled data of the meteorological forecast data for the next 24 hours (the input data and the resampled data have the same format and sampling frequency) into the load forecasting model to obtain the hourly load forecasting values for the next 24 hours.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Any other modifications or equivalent replacements made by those of ordinary skill in the art to the technical solutions of the present invention should be covered within the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A power grid load forecasting method based on multi-source data and hybrid neural network, characterized in that: include: Obtain historical data on the target power grid's load and associated factors; Resampling the historical data according to a preset sampling frequency to obtain resampled data; Perform noise reduction on high-frequency fluctuation data in the resampled data, and perform smoothing on step-change data in the sampled data; Adding time features and weather forecast data to each set of data in the resampled data, and performing normalization processing; Constructing a load forecasting model and using the resampled data to complete the training of the load forecasting model; The load forecasting model is used to forecast the load of the power grid.

2. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 1 is characterized in that: The load forecasting model includes multiple forecasting modules and an output module connected to the output ends of all the forecasting modules; each of the forecasting modules includes an input layer, a CNN module, an LSTM module and an attention mechanism layer connected in sequence; The input layer is used to input the high-frequency fluctuation data; The LSTM module is used to process the step-variation data, the time feature, the weather forecast data and the spliced ​​features of the CNN module output features; The output module is used to perform weighted fusion on the output features of all the prediction modules and output hourly load prediction results.

3. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 2 is characterized in that: Different parameters are used between different CNN modules, different LSTM modules, and different attention mechanism layers.

4. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 2 is characterized in that: At least one of the LSTM modules is provided with all forward LSTM layers, at least one of the LSTM modules is provided with all reverse LSTM layers, and at least one of the LSTM modules is provided with both forward LSTM layers and reverse LSTM layers.

5. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 2 is characterized in that: The output module includes a first fully connected layer, a second fully connected layer and a third fully connected layer; the first fully connected layer is used to dynamically adjust the weights of different prediction modules according to the prediction errors of the prediction modules; The second connection layer is used to perform weighted fusion on the output features of the attention mechanism layer according to the weights of different prediction modules; the third fully connected layer is used to map the fused features of the second connection layer into load prediction results.

6. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 2 is characterized in that: The attention mechanism layer is used to calculate the attention weight for each time step feature output by the LSTM module, and dynamically strengthen the feature expression of the key time step.

7. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 1 is characterized in that: The method for resampling the historical data includes: if the data at the resampling time point in the historical data is missing or abnormal, sampling is performed using an interpolation method; wherein the high-frequency fluctuation data is sampled using a cubic spline interpolation method, and the step-change data is sampled using a forward filling method.

8. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 1 is characterized in that: The autoregressive wavelet denoising method is used to reduce the noise of the high-frequency fluctuation data in the resampled data, and the moving average method is used to smooth the step-change data in the sampled data.

9. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 1, characterized in that: The time features include the annual week number, annual day number, intra-day sampling number, holiday mark and rest day mark corresponding to the sampling time.

10. The power grid load forecasting method based on multi-source data and hybrid neural network according to claim 1, characterized in that: The method for predicting power grid load using a load prediction model includes: obtaining hourly weather forecast data for the next 24 hours and resampling using cubic spline interpolation; inputting sampled data from the previous 48 hours and resampled data of weather forecast data for the next 24 hours into the load prediction model to obtain hourly load prediction values ​​for the next 24 hours.

Images