Comprehensive short-term power load prediction method based on LSTM neural network and CGAN network

Abstract

A comprehensive short-term power load prediction method based on an LSTM neural network and a CGAN network belongs to the short-term power load prediction technology, and comprises the following steps: collecting historical load data of a province and NWP data corresponding to the province, arranging original data, and obtaining historical power consumption data of each city of the province in previous years; calculating weather-load time sequence coupling similarity, sorting the weather-load time sequence coupling similarity according to the similarity, selecting the first n days as a similar day set, and extracting and sorting load data of the first n similar days to form a training set; an LSTM neural network is constructed and trained, a historical 96-point load sequence is used as input, rolling prediction is carried out, and future 96-point loads are predicted; if the prediction day is a non-holiday, the prediction result is a final prediction result; and if the to-be-predicted day is a holiday, performing load correction by using the CGAN network, and adding the correction amount and the initial prediction result to obtain a final prediction result value. By adopting the method, the short-term load prediction precision can be greatly improved.

Description

Technical Field

[0001] This invention pertains to short-term power load forecasting technology, and particularly relates to a comprehensive short-term power load forecasting method based on LSTM neural networks and CGAN networks. Background Technology

[0002] With the deepening of global energy structure transformation and the pursuit of carbon neutrality, the increasing grid connection of new energy sources and the source-load interaction brought about by distributed energy are leading to increasingly complex load dynamics in the power system. Against this backdrop, grid dispatching and operation face higher demands for accurate perception of load changes. Short-term load forecasting, as a key technology supporting the real-time balance and safe operation of the power system, can predict electricity demand hours to days in advance by deeply analyzing historical load data, meteorological factors, and spatiotemporal correlations. Therefore, conducting high-precision short-term load forecasting can not only optimize power generation planning and reduce system operating costs, but also provide data support for grid dispatching, which is of significant practical importance for ensuring the economic efficiency and security of the power system.

[0003] With the development of artificial intelligence technology, its powerful feature extraction and nonlinear fitting capabilities have attracted widespread attention and have been applied in short-term load forecasting. When using neural networks to predict power load, historical data is needed to train the network. However, directly using all historical data for training would result in a large number of irrelevant noise samples that severely interfere with the model's ability to capture true load change patterns, leading to decreased generalization ability and excessive computational resources. Therefore, by using a similar day selection method, historical day data highly similar to the meteorological conditions and load patterns of the forecast day can be accurately selected from massive historical data. This provides a highly correlated sample set for neural network training. This not only greatly improves training efficiency but also allows the neural network to learn more accurate mapping relationships, thereby significantly improving the accuracy and reliability of the prediction model.

[0004] Most current load forecasting methods based on similar days rely primarily on the static similarity of meteorological factors for similarity calculations, meaning they only use meteorological element sequence data to calculate the similarity between the forecast date and historical dates. This method has a significant drawback: it ignores the inherent, strong time periodicity and trend of the load sequence itself. Even existing studies that consider time periodicity only quantify the date as a feature, adding this feature to the meteorological data sequence before selecting similar days. This selection method not only has a vague selection mechanism and poor interpretability, but may also fail to select historical days with the same time period that are more similar to the forecast date due to the influence of other features. For example, a Monday's load curve is not only affected by the weather on that day, but also highly correlated with the trend of the previous week's Monday. Traditional methods might select a Friday with completely identical weather as a similar day for a predicted Monday, but because Friday is at the end of the work cycle and Monday is at the beginning, the load curves of the two are drastically different, leading to a decrease in the final forecast accuracy.

[0005] This method innovatively proposes using a CGAN network for holiday load correction. Existing mainstream holiday correction methods typically fall into two categories: 1) Simple correction method: using a "holiday coefficient" for scaling. This method is too coarse and cannot capture the complex nonlinear load changes under different holiday types and weather conditions. 2) Single model method: attempting to learn weekday and holiday patterns simultaneously using a complex model. However, due to the scarcity and distinct patterns of holiday samples, the model is prone to overfitting to weekday patterns, resulting in poor holiday prediction performance. To address the problem of weak model generalization ability caused by scarce holiday training samples, this method proposes using a CGAN network for holiday load correction. Instead of crudely scaling the data, it transforms the problem into a generative task: allowing the network to learn how to 'correct' from the basic weekday prediction pattern to the actual holiday pattern under constraints such as date type and meteorological characteristics, thus avoiding the model overfitting problem caused by insufficient training samples. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a comprehensive short-term power load forecasting method based on LSTM neural network and CGAN network. By using LSTM neural network and CGAN network in combination, the accuracy of short-term load forecasting is greatly improved.

[0007] The technical solution adopted in this invention is a comprehensive short-term power load forecasting method based on LSTM neural network and CGAN network, comprising the following steps: Step S10: Obtain and organize the historical load data and the corresponding NWP data of a certain province, and at the same time obtain the historical annual electricity consumption data of each city in the province. Step S20: Based on the acquired NWP data and the historical annual electricity consumption data of each city in the province, construct the provincial weighted meteorological data of the province, and align it with the historical load data by time to form a meteorological-load matching dataset; Step S30: Calculate the temporal coupling similarity between meteorological and load data using the meteorological-load matching dataset; The meteorological similarity s1 between the predicted day and the historical day is calculated using provincial weighted meteorological data, and the load similarity s2 between the predicted day and the historical day is calculated using the historical load data of the province. Finally, the meteorological-load time series coupling similarity is obtained. The historical days are sorted from large to small according to the meteorological-load time series coupling similarity, and the top n days are selected as the similar day set. The load data of the top n similar days are extracted and organized to form the training set. Step S40: Construct and train the LSTM neural network; Using the historical load sequence of 96 points as input, rolling forecasts are performed to predict the load output of the next 96 points. If the prediction date is not a public holiday, then the prediction result is the final prediction result. Step S50: Construct and train the CGAN network. If the date to be predicted is a holiday, use the CGAN network to perform load correction. The input to the CGAN network is a feature vector, and the output is the holiday load correction amount; The feature vector consists of meteorological data for the day to be predicted and holiday type data; The correction amount is added to the initial prediction result to obtain the final prediction result value.

[0008] Furthermore, in step 10, the province's historical load data undergoes cleaning, interpolation, and normalization processes.

[0009] Furthermore, NWP data includes data on multiple meteorological elements from various weather stations.

[0010] Further, step S20 includes: Step S201: Based on the historical annual electricity consumption data of each city in the province, calculate the proportion of the historical annual electricity consumption data of each city to the total annual electricity consumption of the province, and use it as the weight of the meteorological element of the city. Step S202: Based on the latitude and longitude coordinates of the meteorological stations in the NWP data, each station is assigned to the corresponding city according to the geographical location of each city in the province, and the average value of the NWP data of all stations in the city is calculated. Step S203: The average value of the NWP data of each city calculated in S202 is weighted and summed according to the weight in step 201 to generate provincial weighted meteorological data. Step S204: Align the provincial weighted meteorological data with the historical load data by time to form a meteorological-load matching dataset.

[0011] Further, step S30 includes: Step S301: Extract the provincial weighted meteorological data and historical load data from the meteorological-load matching dataset, and calculate the correlation coefficient between the provincial weighted data and the historical load data using the Pearson correlation coefficient method. Weights are assigned to each meteorological element based on the absolute value of its correlation coefficient with the load. Step S302: For the day to be predicted and historical days, the correlation coefficient between each meteorological element is calculated using the grey relational analysis method. Based on the weights determined in step S301, the correlation coefficients are weighted and summed to obtain the meteorological similarity s1, which is used to measure the overall similarity between the day to be predicted and historical days in terms of meteorological conditions. Step S303: Extract the daily peak load sequences for the 7 days prior to the forecast date and the 7 days prior to each historical date to construct short-term load change trend characteristics. Again, use the Pearson correlation coefficient method to calculate the correlation coefficient between the two 7-day peak load sequences as the forward trend similarity s2. Step S304: Add the meteorological similarity s1 to the forward trend similarity s2 to obtain the meteorological-load time series coupling similarity S=s1+s2. Sort the meteorological-load time series coupling similarities of all historical days from high to low to form a similarity ranking list. Step S305: Based on the comprehensive similarity ranking results, select the n days with the highest ranking as the similar days; Step S306: Extract load data from similar days and the previous day to form a training set.

[0012] Further, step S50 includes: Step S501: Extract all holiday data from the weather-load matching dataset, and divide the data into training and test sets in an 8:2 ratio; Step S502: For each day in the training set, make predictions according to steps S20-S40, obtain the preliminary prediction results as base data, and subtract the base data from the actual load data to obtain the load correction amount. This load correction amount is the training label of the CGAN model of the model. The training input for the CGAN model is the data sequence of the meteorological element with the highest correlation coefficient calculated in S301 and the input vector consisting of holiday type data; Step S503: Construct the CGAN network. The generator of the CGAN receives the input vector and outputs a 96-point load correction amount. The discriminator determines whether the input conforms to the true correction amount distribution. Both are implemented using an MLP (Multilayer Perceptron). The load correction amount data and the input vector data are used to construct the load correction amount data and train the CGAN network. Step S504: Test the data of the test set. First, predict the date to be predicted in the test set according to steps S20-S40 to obtain the preliminary prediction result. Then, input the input vector data corresponding to the date to be predicted into the CGAN network and output the corresponding load correction amount. The final load prediction result is the sum of the preliminary prediction result and the load correction amount.

[0013] The beneficial effects of this invention are as follows: 1) This invention utilizes a meteorological-load time-series coupling similarity calculation method, which overcomes the shortcomings of traditional methods that only use meteorological data sequences to calculate similarity and thus ignore the time-series characteristics of load sequences; this method can further improve the accuracy of similar day selection, construct higher quality training set data, and thus improve the accuracy of prediction results; 2) A correction method for predicting loads during holidays using CGAN networks is proposed, which effectively avoids the overfitting problem caused by the insufficient number of holiday samples in traditional prediction methods. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the forward trend of the present invention; Figure 3 This is the load curve for different holidays in a certain province. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0016] The main application scenario of this invention is for short-term load forecasting of a province.

[0017] See Figure 1 This invention proposes a comprehensive short-term power load forecasting method based on LSTM neural networks and CGAN networks, comprising the following steps: Step S10: Collect historical load data and corresponding NWP data of a certain province, organize the raw data, and obtain historical electricity consumption data of each city in the province in previous years. Step S20: Using the acquired NWP data and historical electricity consumption data, construct a provincial weighted meteorological data that is more representative of the province's meteorology, and align it with historical load data by time to form a meteorological-load matching dataset; Step S30: Calculate the temporal coupling similarity between meteorological and load data using the meteorological-load matching dataset; The meteorological similarity s1 between the predicted day and the historical day is calculated using provincial weighted meteorological data, and the load similarity s2 between the predicted day and the historical day is calculated using historical load data. Finally, the meteorological-load time series coupling similarity is obtained. The historical days are sorted from large to small according to the meteorological-load time series coupling similarity, and the top n days are selected as the similar day set. The relevant data of the top n similar days are extracted and organized to form the training set. Step S40: Train the LSTM neural network using the data from the constructed training set; Using the historical load sequence of 96 points as input, rolling forecasts are performed to predict the load output of the next 96 points. If the prediction date is not a public holiday, then the prediction result is the final prediction result.

[0018] Step S50: Construct and train the CGAN network; If the forecast date falls on a public holiday, the CGAN network will be used for load correction. The input is a feature vector, and the output is the holiday load correction. The feature vector consists of meteorological data for the day to be predicted and holiday type data; the correction is added to the initial prediction result to obtain the final prediction result.

[0019] In step S10, the historical load data is cleaned, interpolated, and normalized. This method only requires collecting historical load data for a province, without needing to collect historical load data for each city. Simultaneously, it collects historical annual electricity consumption data for each city within the province, which can be obtained from local government statistical platforms using online search functions.

[0020] The acquired historical load data is cleaned to remove obviously abnormal or duplicate records, and missing parts are filled using time series interpolation or regional mean imputation methods.

[0021] Step S20 includes: Step S201: Based on the historical electricity consumption data of each city, calculate the proportion of each city's electricity consumption to the province's total, and use this as the weight of the meteorological elements of that city. Specifically, assuming the province has n prefecture-level cities, and the collected historical electricity consumption data for each city are: E1, E2, E3, ..., En, then the weight of each city is calculated using the following formula: , Step S202: Based on the latitude and longitude coordinates of the meteorological stations in the NWP data, each station is assigned to the corresponding city according to the geographical location of each city in the province, and the average value of the NWP data of all stations in the city is calculated. Assuming the city's geographical location on the map is between 102°35′58” and 104°34′29” east longitude and between 35°34′20” and 37°07′07” north latitude, based on the latitude and longitude information of meteorological stations in the NWP data, stations within this range are included in the city's scope, and these stations are assigned the corresponding weights to the city when calculating weighted meteorological data in the subsequent calculation.

[0022] Step S203: The average value of the NWP data of each city calculated in S202 is weighted and summed according to the weight in step 201 to generate provincial weighted meteorological data. Suppose the province has n cities, and each city has If there are q meteorological stations, the q-th meteorological indicator of the province is calculated using the following formula: , The formula is used to calculate the provincial weighted meteorological data for each meteorological element, and the resolution of the data is consistent with that of the NWP data.

[0023] Step S204 involves aligning the weighted meteorological data with historical load data by time to form a meteorological-load matching dataset for subsequent similarity calculations. Note that this step requires maintaining the same data resolution between the load data and the meteorological data; all data should be processed to a 15-minute resolution.

[0024] Step S30 includes: Step S301: Extract the provincial weighted meteorological data and historical load data from the meteorological-load matching dataset, and calculate the correlation coefficient between the provincial weighted data and the historical load data using the Pearson correlation coefficient method; Weights are assigned to each meteorological element based on the absolute value of its correlation coefficient with the load. Assumption These are the q meteorological element sequences of the province. L is the corresponding load element sequence The Pearson correlation coefficient method was used to calculate the Pearson correlation coefficient between each meteorological element and the load series. As shown in the following formula: , In the formula, n is the sequence length.

[0025] Based on the calculated correlation coefficients, each meteorological element is assigned a weight using the following formula; ; Step S302: For the day to be predicted and historical days, the correlation coefficient between each meteorological element is calculated using the grey relational analysis method. According to the weights determined in step S301, the correlation coefficients are weighted and summed to obtain the meteorological similarity s1, which is used to measure the overall similarity between the day to be predicted and historical days in terms of meteorological conditions. Assume the sequence of all meteorological elements for the day to be predicted is as follows: , Let represent the sequence of meteorological elements for the day to be predicted; the same logic applies to the remaining meteorological elements, assuming each meteorological element has 96 data points. The meteorological element sequence for the i-th historical day is... Where D represents the data for the day to be predicted.

[0026] The correlation coefficients between meteorological elements on the day to be predicted and the i-th historical day are calculated using the following formula: , in The resolution coefficient, with a value between 0 and 1, is used in this invention. Take 0.5; For the predicted day and the i-th historical day The similarity of individual meteorological elements; For the m-th meteorological element of the day to be predicted, the k-th data point is provided. This represents the k-th data point of the m-th meteorological element on the i-th historical day.

[0027] The meteorological similarity between the predicted date and historical dates is calculated using the following formula: , The meteorological similarity s1 is obtained, which is used to measure the overall similarity of the meteorological conditions between the day to be predicted and the historical days.

[0028] Step S303: Extract the daily peak load sequences for the 7 days prior to the forecast date and the 7 days prior to each historical date to construct short-term load change trend characteristics. Again, use the Pearson correlation coefficient method to calculate the correlation coefficient between the two 7-day peak load sequences as the forward trend similarity s2. Assume the load peak sequence for each of the seven days preceding the forecast date is as follows: ,in The peak load for the day before the forecast date. This represents the peak load of the previous day. The load peak value is the 7th consecutive day prior to the i-th historical day; the load peak value sequence for each of the seven days preceding the i-th historical day is as follows: The correlation between two sequences is calculated using the Pearson correlation coefficient method to obtain the pre-trend similarity s2; a schematic diagram of the pre-trend similarity is shown below. Figure 2 As shown; Step S304: Add the meteorological similarity s1 to the forward trend similarity s2 to obtain the comprehensive similarity. The similarity of the weather-load time series coupling of all historical days is sorted from high to low to form a similarity ranking list; Step S305: Based on the similarity ranking results, select the top 125 days as the similarity days. This number can be adjusted according to the scale of historical data and prediction needs, aiming to balance the computational complexity of model training with prediction accuracy.

[0029] Step S306: Extract data from similar days and the previous day to form a training set. The training set includes input features: a historical load sequence of 96 points and output labels; and the load value at the time point to be predicted. By constructing the training set in this way, 96 training samples can be obtained each day. It is important to note that holiday data should not be included in the historical days used to calculate similarity; only non-holiday data should be used for selecting similar days. Step S50 includes: Step S501: Extract all holiday data from the historical data, and divide the data into training and test sets in an 8:2 ratio; The specific holiday data includes: New Year's Day, Spring Festival, Qingming Festival, Labor Day, Dragon Boat Festival, Mid-Autumn Festival, and National Day.

[0030] Step S502: Construct the training set data for CGAN: Use the load prediction method described above to predict the data in the training set and obtain its preliminary prediction results as base data; The load correction is obtained by subtracting the base data from the actual load forecast data. This load correction is the training label of the CGAN model. The feature vector input to the model consists of: the meteorological data sequence for the day to be predicted and holiday type data. Where H represents the holiday type data for the date to be predicted, and the corresponding values ​​are shown in the table below: Holidays New Year Spring Festival Qingming Festival Labor Day Dragon Boat Festival Mid-Autumn Festival National Day Data Tags 1 2 3 4 5 6 7 It is noted that different types of holidays exhibit significantly different load trends due to variations in social customs and holiday duration. Therefore, to improve model accuracy, specific categories of holidays are included in the model input to capture their internally differentiated time-series features. For example... Figure 3 The figures show the holiday load curves for New Year's Day, Qingming Festival, Labor Day, and Dragon Boat Festival, respectively. It can be seen that different holiday types have significantly different load trends.

[0031] Step S503: Construct the CGAN network; The CGAN generator receives the input vector and outputs a 96-point load correction; the discriminator determines whether the input comes from the true correction distribution. In this invention, both are implemented using a multilayer perceptron. Using the constructed training set data, CGAN is adversarially trained. By minimizing the adversarial loss between the generator and the discriminator, the generator can output results that conform to the true correction distribution.

[0032] Step S504: Finally, test the test set data: First, use the LSTM neural network based on similar days to make a preliminary prediction and obtain a preliminary prediction result. Second, input the input vector data corresponding to the day to be predicted into the CGAN network and output the corresponding load correction amount. The final load prediction result is the sum of the preliminary prediction result and the load correction amount.

[0033] 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; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A comprehensive short-term power load forecasting method based on LSTM neural network and CGAN network, characterized in that, Includes the following steps: Step S10: Obtain and organize the historical load data and the corresponding NWP data of a certain province, and at the same time obtain the historical annual electricity consumption data of each city in the province. Step S20: Based on the acquired NWP data and the historical annual electricity consumption data of each city in the province, construct the provincial weighted meteorological data of the province, and align it with the historical load data by time to form a meteorological-load matching dataset; Step S30: Calculate the temporal coupling similarity between meteorological and load data using the meteorological-load matching dataset; The meteorological similarity s1 between the predicted day and the historical day is calculated using provincial weighted meteorological data, and the load similarity s2 between the predicted day and the historical day is calculated using the historical load data of the province. Finally, the meteorological-load time series coupling similarity is obtained. The historical days are sorted from large to small according to the meteorological-load time series coupling similarity, and the top n days are selected as the similar day set. The load data of the top n similar days and the day before the similar days are extracted and organized to form the training set. Step S40: Construct and train the LSTM neural network; Using the historical load sequence of 96 points as input, rolling forecasts are performed to predict the load output of the next 96 points. If the prediction date is not a public holiday, then the prediction result is the final prediction result. Step S50: Construct and train the CGAN network. If the date to be predicted is a holiday, use the CGAN network to perform load correction. The input to the CGAN network is a feature vector, and the output is the holiday load correction amount; The feature vector consists of meteorological data for the day to be predicted and holiday type data; The correction amount is added to the initial prediction result to obtain the final prediction result value.

2. The comprehensive short-term power load forecasting method according to claim 1, characterized in that, In step 10, the province's historical load data is cleaned, interpolated, and normalized.

3. The comprehensive short-term power load forecasting method according to claim 1, characterized in that, NWP data contains data on multiple meteorological elements from various weather stations.

4. The comprehensive short-term power load forecasting method according to claim 1, characterized in that, Step S20 includes: Step S201: Based on the historical annual electricity consumption data of each city in the province, calculate the proportion of the historical annual electricity consumption data of each city to the total annual electricity consumption of the province, and use it as the weight of the meteorological element of the city. Step S202: Based on the latitude and longitude coordinates of the meteorological stations in the NWP data, each station is assigned to the corresponding city according to the geographical location of each city in the province, and the average value of the NWP data of all stations in the city is calculated. Step S203: The average value of the NWP data of each city calculated in S202 is weighted and summed according to the weight in step 201 to generate provincial weighted meteorological data. Step S204: Align the provincial weighted meteorological data with the historical load data by time to form a meteorological-load matching dataset.

5. The comprehensive short-term power load forecasting method according to claim 1, characterized in that, Step S30 includes: Step S301: Extract the provincial weighted meteorological data and historical load data from the meteorological-load matching dataset, and calculate the correlation coefficient between the provincial weighted data and the historical load data using the Pearson correlation coefficient method. Weights are assigned to each meteorological element based on the absolute value of its correlation coefficient with the load. Step S302: For the day to be predicted and historical days, the correlation coefficient between each meteorological element is calculated using the grey relational analysis method. Based on the weights determined in step S301, the correlation coefficients are weighted and summed to obtain the meteorological similarity s1, which is used to measure the overall similarity between the day to be predicted and historical days in terms of meteorological conditions. Step S303: Extract the daily peak load sequences for the 7 days prior to the forecast date and the 7 days prior to each historical date to construct short-term load change trend characteristics. Again, use the Pearson correlation coefficient method to calculate the correlation coefficient between the two 7-day peak load sequences as the forward trend similarity s2. Step S304: Add the meteorological similarity s1 to the forward trend similarity s2 to obtain the meteorological-load time series coupling similarity S=s1+s2. Sort the meteorological-load time series coupling similarities of all historical days from high to low to form a similarity ranking list. Step S305: Based on the comprehensive similarity ranking results, select the n days with the highest ranking as the similar days; Step S306: Extract load data from similar days and the previous day to form a training set.

6. The integrated short-term power load forecasting method based on LSTM neural network and CGAN network according to claim 1, characterized in that, Step S50 includes: Step S501: Extract all holiday data from the weather-load matching dataset, and divide the data into training and test sets in an 8:2 ratio; Step S502: For each day in the training set, make predictions according to steps S20-S40, obtain the preliminary prediction results as base data, and subtract the base data from the actual load data to obtain the load correction amount. This load correction amount is the training label of the CGAN model of the model. The training input for the CGAN model is the data sequence of the meteorological element with the highest correlation coefficient calculated in S301 and the input vector consisting of holiday type data; Step S503: Construct the CGAN network. The generator of the CGAN receives the input vector and outputs a 96-point load correction amount. The discriminator determines whether the input conforms to the true correction amount distribution. Both are implemented using an MLP (Multilayer Perceptron). The load correction amount data and the input vector data are used to construct the load correction amount data and train the CGAN network. Step S504: Test the data of the test set. First, predict the date to be predicted in the test set according to steps S20-S40 to obtain the preliminary prediction result. Then, input the input vector data corresponding to the date to be predicted into the CGAN network and output the corresponding load correction amount. The final load prediction result is the sum of the preliminary prediction result and the load correction amount.

Images