Load prediction method, system and equipment based on time sequence characteristics and medium

Abstract

The invention discloses a time sequence characteristic-based load prediction method, system and device and a medium, and the method employs an LSTM and Attention mechanism to construct a time sequence residual neural network, can accurately capture the time sequence change rule of load data in different scenes, solves a defect that an XGBoost model is weak in time sequence characteristic capture, and improves the load prediction efficiency. According to the invention, a weighted loss function based on a sample time sequence is introduced, the attention of the model to recent samples is increased, the load prediction accuracy is improved, the challenge of new energy power generation influence and meteorological factor coupling on load change prediction is effectively solved, and the load prediction efficiency is improved. And meanwhile, the demand for additional characteristic variables is reduced, the calculation consumption in the model training and prediction process is reduced, the method better adapts to the time sequence rule change of the load and the coupling of related elements, and the load prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to a load forecasting technology for an electric power system, and in particular to a load forecasting method, system, equipment and medium based on time series characteristics. Background Art

[0002] Accurate load forecasting is the basis for optimizing the allocation and dispatching of power system resources. It can better coordinate the grid connection of renewable energy, reduce the phenomenon of wind and solar power abandonment, and improve the utilization rate of renewable energy. In addition, the load forecasting results can be used to guide the charging and discharging strategies of energy storage systems, realize efficient energy storage and release, and realize the reasonable optimization of dispatching and resource allocation, further improving the flexibility and reliability of the power grid. The precision and stabilization of load forecasting will further improve the automation level of the power grid dispatching system and improve the degree of economic dispatching; it will become an important basis for the power spot market system to formulate clearing plans, affecting the reliability of clearing plans; it will become a solid foundation for the new generation of power grid intelligent dispatching systems, providing reliable load demand information for upper-level decision-making.

[0003] At present, there are two main development directions for load forecasting methods: one is the time series neural network route, which is mainly based on the historical load data, combined with load-related meteorological, holidays and other information to fit the load time series changes, but with the increasing impact of current new energy power generation, the change of load is more obviously affected by meteorological factors, and the coupling between meteorological elements makes the load show different change rules, that is, multiple scenes, and the load samples in each scene are small, so the time series neural network cannot fit the load change rules in multiple scenes well, affecting the accuracy of the model. The second is the tree model architecture represented by XGBoost. The main considerations are similar to the neural network route, but because the tree model architecture is difficult to consider the time series changes of the load and the coupling of various related elements, it is necessary to add additional related feature variables to improve the model's capture of the load law, which brings additional computing consumption, and it is also difficult to capture the time series variability of the load. For example, with the development of the economy, the load has continued to grow and set a new high every year, and the model's fitting ability is poor. Summary of the invention

[0004] Purpose of the invention: The present invention aims to provide a load forecasting method that uses a time series residual neural network to capture load time series characteristics, solves the defect of the XGBoost model in capturing time series characteristics weakly, and alleviates the lag of traditional LSTM time series neural network in fitting load changes in multiple scenarios; another purpose of the present invention is to provide a load forecasting system, equipment and medium based on time series characteristics.

[0005] Technical solution: The load forecasting method based on time series characteristics described in the present invention comprises the following steps:

[0006] Align the regional load historical data and historical meteorological data in time series;

[0007] After cleaning the time-series aligned regional load historical data and historical meteorological data, the historical load data and historical meteorological data are spliced ​​in time series to construct a load forecast historical sample set;

[0008] According to the historical load changes, the basic features of the historical sample set are constructed to obtain the basic sample set;

[0009] Preprocess the basic sample set, use LSTM and Attention mechanism to build a time series residual neural network, train and screen the time series residual neural network through the preprocessed basic sample set, use the screened time series residual neural network to extract time series features from the basic sample set, and obtain the time series sample set;

[0010] The time series sample set and the basic sample set are concatenated to obtain the load forecasting feature data set. The load forecasting feature data set is used to train the XGBoost model to obtain the optimal model for load forecasting.

[0011] Furthermore, the basic features of the historical sample set are constructed according to the historical load changes to obtain the basic sample set, as follows:

[0012] Construct corresponding load fluctuation cycle characteristics according to daily or weekly load fluctuations;

[0013] Construct scenario-differentiated onehot encoding features based on load scenario differences.

[0014] Furthermore, the basic sample set is preprocessed, and the time series residual neural network is constructed using the LSTM and Attention mechanisms. The time series residual neural network is trained and screened using the preprocessed basic sample set, and the screened time series residual neural network is used to extract time series features from the basic sample set to obtain a time series sample set, as follows:

[0015] After normalizing the basic sample set, a periodic analysis is performed, and based on the periodic analysis, the basic sample set is sampled at a minimum period to obtain a basic data set;

[0016] Split the basic data set to obtain a basic training set and a basic validation set;

[0017] Based on the basic training set, the LSTM and attention time series residual neural network are trained, and the basic validation set is used to screen the model;

[0018] A time series residual neural network is used to extract time series features from the basic sample set, and the extracted features are time-series spliced ​​with the basic sample set to obtain a time series sample set.

[0019] Furthermore, a periodic analysis is performed on the basic sample set. According to the periodic analysis, the basic sample set is sampled at a minimum period to obtain a basic data set, as follows:

[0020] Wavelet transform is used to extract the load history data cycle, and the layer with the shortest cycle is found from all decomposition levels, which is defined as the minimum load cycle;

[0021] The basic data set is uniformly sampled according to the minimum load period to obtain the basic data set.

[0022] Furthermore, based on the basic training set, the time series residual neural network constructed using LSTM and attention is trained, and the model is screened using the basic validation set, as follows:

[0023] The load time series feature extraction block is constructed through the embedding layer, Attention, LSTM and residual connection to obtain the load time series features;

[0024] The meteorological time series feature extraction block is constructed through the embedding layer, Attention, bidirectional LSTM and residual connection to obtain the meteorological time series features;

[0025] The load time series feature extraction features are spliced ​​with the meteorological time series features to fit the load data of the next cycle;

[0026] Adjust the input sequence length to an integer multiple of the minimum period based on the basic data set, and save the time series feature extraction model obtained with different input sequence lengths;

[0027] The degree of fit of each time series feature extraction model is evaluated based on the validation set to determine the optimal time series feature extraction model.

[0028] Furthermore, the optimal time series feature extraction model is used to extract time series features from the basic sample set, and the extracted features are arranged in time series to obtain a time series sample set.

[0029] Furthermore, the time series sample set and the basic sample set are concatenated to obtain a load forecast feature data set, and the load forecast feature data set is used to train the XGBoost model to obtain the optimal model for load forecasting, as follows:

[0030] The time series sample set and the basic sample set are spliced ​​in time series to obtain the load forecasting feature data set.

[0031] The load forecast feature data set is uniformly sampled into training set and validation set in proportion, the time sequence of the data set is retained, and the most recent data is divided into the forecast set according to the time sequence;

[0032] Based on the temporal nature of the feature data set, a loss function WLoss based on sample temporal weighting is introduced:

[0033]

[0034] Among them, y i is the true load of the ith sample, is the predicted load of the i-th sample, N is the total number of samples, ω i is the weight of the i-th sample, ω i for

[0035] ω i =exp(-α·(Ni))

[0036] Among them, α is a constant that controls the decay rate;

[0037] The XGBoost model is trained using the training set, and the model is screened based on its performance in the validation set to obtain the optimal model for load forecasting.

[0038] The load forecasting system based on time series characteristics of the present invention comprises:

[0039] Data acquisition and alignment module, used to acquire historical regional load data and historical meteorological data, and align the data in time sequence;

[0040] The data cleaning and splicing module is used to clean the historical load data and historical meteorological data, and splice the historical load data and historical meteorological data in time series to construct a historical sample set for load forecasting;

[0041] A basic feature construction module is used to construct basic features of the historical sample set according to the historical load changes to obtain a basic sample set;

[0042] The time series feature extraction module is used to construct a time series residual neural network based on the basic sample set using LSTM and Attention mechanisms to extract the time series features of the basic data and obtain the time series sample set;

[0043] The model verification and prediction module is used to concatenate the time series sample set and the basic sample set to obtain a load forecast feature data set, and use the load forecast feature data set to train the XGBoost model to obtain the optimal model for load forecasting.

[0044] The computer device of the present invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the steps of the above method are implemented when the processor executes the computer program.

[0045] The computer-readable storage medium of the present invention stores a computer program thereon, and the computer program implements the steps of the above method when executed by a processor.

[0046] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The load forecasting method based on time series characteristics of the present invention utilizes LSTM and Attention mechanism to construct a time series residual neural network, which can accurately capture the time series change law of load data in different scenarios, solves the defect of XGBoost model in weak capture of time series features, alleviates the lag of traditional LSTM time series neural network in fitting load changes in multiple scenarios, and further introduces a weighted loss function based on sample time series, increases the model's attention to recent samples, improves the load forecasting accuracy, and effectively solves the challenges of the coupling of new energy power generation and meteorological factors to load change prediction; 2. The load forecasting method based on time series characteristics of the present invention reduces the demand for additional feature variables and reduces the computational consumption in the model training and prediction process; 3. Compared with the traditional tree model, the load forecasting method based on time series characteristics of the present invention can better adapt to the time series law changes of load and the coupling of related factors, and improves the load forecasting accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of the flow of the load forecasting method based on time series characteristics of the present invention;

[0048] Figure 2 Extract neural network architecture diagram for time series features;

[0049] Figure 3 A diagram showing the prediction results of the load prediction method based on time series characteristics described in the present invention;

[0050] Figure 4 This is a comparison chart of the prediction results of the load prediction method based on time series characteristics described in the present invention and other models. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings.

[0052] The load forecasting method based on time series characteristics of the present invention comprises the following steps:

[0053] S1. Obtain historical regional load data and historical meteorological data, and align the data in time series.

[0054] In this embodiment, numerical weather forecast meteorological data and load history data of a province from 2021-1-1 to 2024-10-30 were obtained. The relevant weather data included: temperature, rainfall, irradiance, wind speed, humidity, air pressure, etc.

[0055] S2. Clean the historical load data and historical meteorological data, and splice the historical load data and historical meteorological data in time series to construct a load forecast historical sample set, which specifically includes the following steps:

[0056] S2.1. Detect missing values, zero values, and sudden change values ​​in load historical data and numerical weather forecast meteorological data, and set the detected abnormal values ​​to null.

[0057] S2.2. Use bilinear interpolation, spline interpolation, mean interpolation and other methods to fill in the empty values ​​to ensure the time series integrity of load data and weather data.

[0058] S3. Construct basic features of the historical sample set according to the historical load changes to obtain a basic sample set, which specifically includes the following steps:

[0059] S3.1. Based on the daily load volatility, the daily cycle characteristics of load fluctuation are constructed, and the daily cycle is periodically encoded using sin and cos functions to capture the daily periodic changes of load;

[0060] S3.2. According to the weekly load fluctuation, the load fluctuation cycle characteristics are constructed, and the sin and cos functions are used to encode the weekly cycle to capture the weekly periodic changes of the load;

[0061] S3.2. Construct scene-differentiated onehot coding features according to load scene differences, and perform onehot coding for different holidays, weekends, continuous high temperature, continuous low temperature, and high and low temperature sudden change scenes;

[0062] S4. Based on the basic sample set, the LSTM and Attention mechanisms are used to construct a time series residual neural network to extract the time series features of the basic data and obtain a time series sample set, which specifically includes:

[0063] S4.1. Normalize the basic sample set to eliminate the data distribution deviation caused by different dimensions. The normalization calculation formula is:

[0064]

[0065] Where X is the basic sample set, X∈R NxM , where N represents the number of samples and M represents the basic feature dimension.

[0066] S4.2. Perform periodic analysis on the basic sample set. According to the periodic analysis, sample the basic sample set at the minimum period to obtain the basic data set, which specifically includes:

[0067] Wavelet transform is used to extract the load history data cycle, and the layer with the shortest cycle is found from all decomposition levels, which is the minimum load cycle;

[0068] The basic data set is uniformly sampled according to the minimum load period to obtain the basic data set.

[0069] In this step, wavelet transform is used to extract the load history data cycle, and the layer with the shortest cycle is found from all decomposition levels, which is the minimum load cycle, including the following steps:

[0070] S4.2.1. Based on historical load data, select the wavelet transform type and decomposition level;

[0071] S4.2.2, perform multi-layer wavelet decomposition on historical load data;

[0072] S4.2.3, estimate the period based on the number of decomposition layers and the corresponding frequency bands of each layer;

[0073] S4.2.4. Find the layer with the shortest period from all decomposition levels, which is the minimum period of the load sequence;

[0074] S4.3, split the basic data set to obtain a basic training set and a basic verification set;

[0075] S4.4. Based on the basic training set, use LSTM and attention to build a time series residual neural network to train the load, and use the basic validation set to screen the model, including:

[0076] S4.4.1. The load time series feature extraction block is constructed through the embedding layer, Attention, LSTM and residual connection to obtain the load time series features, including:

[0077] 1) Use the embedding layer network and attention mechanism to enhance the ability to express historical load data and calculate the impact relationship of different time steps. The calculation steps can be expressed as:

[0078]

[0079] in:

[0080] embed(x) = ReLU(W e1 x+b e1 )W e2 +b e2

[0081] Among them, x1 is the load input sequence, W e1 , W e2 、b e1 、b e2 All are learnable parameters, and attention is the attention score calculation.

[0082] 2) Use LSTM to capture the timing characteristics of the load sequence and obtain the timing representation of the load input sequence. The calculation steps can be expressed as:

[0083]

[0084] where h t is the time series representation of the load input sequence at the tth moment, c t represents the potential memory update at time step t.

[0085] S4.4.2, construct a meteorological time series feature extraction block through embedding layer, Attention, bidirectional LSTM and residual connection to obtain meteorological time series features. The specific calculation method is similar to S4.4.1;

[0086] S4.4.3. Combine the load time series feature extraction features with the meteorological time series features to fit the load data of the next cycle. The calculation steps can be expressed as:

[0087] output = tanh(W2σ(W1concat(h load,t ,h weather,t )+b1)+b2)

[0088] where h load,t 、h weather,t are the time series representation of the load input sequence at the tth moment and the time series representation of the meteorological input sequence at the tth moment respectively. W1, W2, b1, and b2 are all learnable parameters.

[0089] S4.4.4. Adjust the input sequence length based on the basic data set to make it an integer multiple of the minimum period, and save the time series feature extraction model weights obtained with different input sequence lengths;

[0090] S4.4.5. Evaluate the degree of fit of each time series feature extraction model based on the validation set, select the best time series feature extraction model, and save the model weights.

[0091] S4.5. Use the time series residual neural network to extract time series features from the entire basic sample set, and perform time series splicing with the basic sample set to obtain a time series sample set, specifically including:

[0092] S4.5.1. Based on the predicted output sequence length, the output layer of the feature extraction model is transformed in dimension to adapt it to the output sequence length;

[0093] S4.5.2. Load the time series feature extraction model weights and fine-tune the time series feature extraction model output layer weights based on the basic sample set;

[0094] S4.5.3. Use the fine-tuned time series feature extraction model to extract time series features from the basic sample set, and arrange the extracted features in time series to obtain a time series sample set.

[0095] S5. The time series sample set and the basic sample set are concatenated to obtain a load forecast feature data set, and the feature data set is further used to train, verify and predict the feature data set using an XGBoost model.

[0096] S5.1. splicing the time series sample set and the basic sample set in time series to obtain a load forecasting characteristic data set;

[0097] S5.2, divide the feature data set, sample it evenly in proportion into a training set and a validation set, retain the time sequence of the data set, and divide the most recent data into a prediction set according to the time sequence;

[0098] S5.3. Based on the temporal nature of the feature data set, a loss function based on sample temporal weighting is introduced to make the model pay more attention to recent data errors. The loss function calculation formula is as follows:

[0099]

[0100] where y i is the true load of the ith sample, is the predicted load of the i-th sample, N is the total number of samples, ω i is the weight of the i-th sample, which is calculated as:

[0101] ω i =exp(-α·(Ni))

[0102] Here, α is a constant that controls the decay rate, and its value is 0.05 in this example.

[0103] S5.4. Use the training set to train the XGBoost model, and screen the model based on its performance in the validation set to obtain a prediction model;

[0104] S5.5. Make predictions for the prediction set based on the prediction model, compare the predicted values ​​with the actual values, and evaluate the model performance.

[0105] In this embodiment, the period from 2021-1-1 to 2024-9-10 in a certain province is used as the training set and validation set, and the period from 2024-9-11 to 2024-9-29 is used as the prediction set. Finally, the model is evaluated by comparing the difference between the measured values ​​and the predicted values ​​in the prediction set.

[0106] The load forecasting system based on time series characteristics of the present invention can realize high-precision load forecasting, and specifically includes:

[0107] Data acquisition and alignment module, used to acquire historical regional load data and historical meteorological data, and align the data in time sequence;

[0108] The data cleaning and splicing module is used to clean the historical load data and historical meteorological data, and splice the historical load data and historical meteorological data in time series to construct a historical sample set for load forecasting;

[0109] A basic feature construction module is used to construct basic features of the historical sample set according to the historical load changes to obtain a basic sample set;

[0110] The time series feature extraction module is used to construct a time series residual neural network based on the basic sample set using LSTM and Attention mechanisms to extract the time series features of the basic data and obtain the time series sample set;

[0111] The model verification and prediction module is used to concatenate the time series sample set and the basic sample set to obtain a load forecast feature data set, and use the load forecast feature data set to train the XGBoost model to obtain the optimal model for load forecasting.

[0112] The computer device of the present invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the steps of the above method are implemented when the processor executes the computer program.

[0113] The computer-readable storage medium of the present invention stores a computer program thereon, and the computer program implements the steps of the above method when executed by a processor.

Claims

1. A load forecasting method based on time series characteristics, characterized in that: The following steps are involved: Align the regional load historical data and historical meteorological data in time series; After cleaning the time-series aligned regional load historical data and historical meteorological data, the historical load data and historical meteorological data are spliced ​​in time series to construct a load forecast historical sample set; According to the historical load changes, the basic features of the historical sample set are constructed to obtain the basic sample set; Preprocess the basic sample set, use LSTM and Attention mechanism to build a time series residual neural network, train and screen the time series residual neural network through the preprocessed basic sample set, use the screened time series residual neural network to extract time series features from the basic sample set, and obtain the time series sample set; The time series sample set and the basic sample set are concatenated to obtain the load forecasting feature data set. The load forecasting feature data set is used to train the XGBoost model to obtain the optimal model for load forecasting.

2. The load forecasting method based on time series characteristics according to claim 1 is characterized in that: According to the historical load changes, the basic features of the historical sample set are constructed to obtain the basic sample set, as follows: Construct corresponding load fluctuation cycle characteristics according to daily or weekly load fluctuations; Construct scenario-differentiated onehot encoding features based on load scenario differences.

3. The load forecasting method based on time series characteristics according to claim 1 is characterized in that: The basic sample set is preprocessed, and the time series residual neural network is constructed using the LSTM and Attention mechanisms. The time series residual neural network is trained and screened through the preprocessed basic sample set. The screened time series residual neural network is used to extract time series features from the basic sample set to obtain the time series sample set, as follows: After normalizing the basic sample set, a periodic analysis is performed, and based on the periodic analysis, the basic sample set is sampled at a minimum period to obtain a basic data set; Segment the basic data set; Get the basic training set and basic validation set; Based on the basic training set, the LSTM and attention time series residual neural network are trained, and the basic validation set is used to screen the model; The time series residual neural network is used to extract the time series features of the basic sample set to obtain the time series sample set, and the time series sample set is spliced ​​with the basic sample set to obtain the load forecasting feature data set.

4. The load forecasting method based on time series characteristics according to claim 3 is characterized in that: Perform periodic analysis on the basic sample set, and obtain the basic data set by sampling the basic sample set at the minimum period according to the periodic analysis, as follows: Wavelet transform is used to extract the load history data cycle, and the layer with the shortest cycle is found from all decomposition levels, which is defined as the minimum load cycle; The basic data set is uniformly sampled according to the minimum load period to obtain the basic data set.

5. The load forecasting method based on time series characteristics according to claim 3 is characterized in that: Based on the basic training set, the time series residual neural network constructed using LSTM and attention is trained, and the model is screened using the basic validation set, as follows: The load time series feature extraction block is constructed through the embedding layer, Attention, LSTM and residual connection to obtain the load time series features; The meteorological time series feature extraction block is constructed through the embedding layer, Attention, bidirectional LSTM and residual connection to obtain the meteorological time series features; The load time series feature extraction features are spliced ​​with the meteorological time series features to fit the load data of the next cycle; Adjust the input sequence length to an integer multiple of the minimum period based on the basic data set, and save the time series feature extraction model obtained with different input sequence lengths; The degree of fit of each time series feature extraction model is evaluated based on the validation set to determine the optimal time series feature extraction model.

6. The load forecasting method based on time series characteristics according to claim 5 is characterized in that: The best optimal time series feature extraction model is used to extract time series features from the basic sample set, and the extracted features are arranged in time series to obtain a time series sample set.

7. The load forecasting method based on time series characteristics according to claim 6 is characterized in that: The time series sample set and the basic sample set are concatenated to obtain the load forecast feature data set. The load forecast feature data set is used to train the XGBoost model to obtain the optimal model for load forecasting, as follows: The time series sample set and the basic sample set are spliced ​​in time series to obtain the load forecasting feature data set. The load forecast feature data set is uniformly sampled into training set and validation set in proportion, the time sequence of the data set is retained, and the most recent data is divided into the forecast set according to the time sequence; Based on the temporal nature of the feature data set, a loss function WLoss based on sample temporal weighting is introduced: Among them, y i is the true load of the ith sample, is the predicted load of the i-th sample, N is the total number of samples, ω i is the weight of the i-th sample, ω i for oh i =exp(-α·(Ni)) Among them, α is a constant that controls the decay rate; The XGBoost model is trained using the training set, and the model is screened based on its performance in the validation set to obtain the optimal model for load forecasting.

8. A load forecasting system based on time series characteristics, characterized in that: include: Data acquisition and alignment module, used to acquire historical regional load data and historical meteorological data, and align the data in time sequence; The data cleaning and splicing module is used to clean the historical load data and historical meteorological data, and splice the historical load data and historical meteorological data in time series to construct a historical sample set for load forecasting; A basic feature construction module is used to construct basic features of the historical sample set according to the historical load changes to obtain a basic sample set; The time series feature extraction module is used to construct a time series residual neural network based on the basic sample set using LSTM and Attention mechanisms to extract the time series features of the basic data and obtain the time series sample set; The model verification and prediction module is used to concatenate the time series sample set and the basic sample set to obtain a load forecast feature data set, and use the load forecast feature data set to train the XGBoost model to obtain the optimal model for load forecasting.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images