Daily electricity consumption recursive multi-step prediction method and device considering weather similar days

Abstract

The invention discloses a daily electricity consumption recursive multi-step prediction method and device considering similar weather days, and belongs to the daily electricity consumption prediction technology, and the method comprises the steps: obtaining historical day meteorological element data and electricity consumption data of a target region, and screening key meteorological elements; calculating comprehensive similarity by using the key meteorological elements, and selecting meteorological similar days; extracting and sorting data of similar days and previous days to form training set data; training a neural network; and using the trained neural network, taking the data of the first 24 hours of the prediction day as input, and performing recursive prediction to obtain a prediction result of the power consumption of the prediction day. The device is used for implementing the method. According to the method, the similar day algorithm and the recursion multi-step prediction method are organically combined, the pertinence of the model to different weather conditions is improved by using the similar day algorithm, the time sequence characteristics of the electricity consumption data are mined by using the recursion multi-step prediction method, and the precision of the daily electricity consumption prediction result is improved.

Description

Technical Field

[0001] The present invention belongs to the technology of daily power consumption prediction, and particularly relates to a recursive multi-step prediction method and device for daily power consumption considering meteorologically similar days. Background Art

[0002] In the power system, due to the limitation of electrical energy storage, real-time supply-demand balance is particularly crucial. Especially in a complex power market environment and changing load demands, accurate power consumption prediction is of great importance.

[0003] Daily power consumption is closely related to the meteorological conditions of the day, and the corresponding power consumption patterns under different meteorological conditions have significant differences. The power consumption prediction methods have been continuously developed, and currently, the similar-day method is mainly adopted.

[0004] Existing power consumption prediction methods usually directly use all historical meteorological and power consumption data to train the same model. For example, a short-term load prediction method disclosed in Chinese Patent Application CN109636063A is difficult to comprehensively cover diverse meteorological scenarios and lacks pertinence, resulting in poor scenario adaptability of the model.

[0005] Traditional similar-day selection algorithms mainly rely on calculating the correlation of multiple meteorological factors between historical days and prediction days and taking their average value to evaluate similarity. For example, a short-term power load two-way combined prediction method based on similar days and LSTM proposed in Chinese Patent CN111932402B fails to fully reflect the differences in the influence degrees of various meteorological factors on power consumption and may mask key factors due to the high correlation of non-critical factors, resulting in high similarity but inconsistent power consumption patterns.

[0006] In addition, existing technologies mainly use the daily total power consumption data of meteorologically similar days for prediction. This method ignores the fluctuations in power consumption within a day, cannot reflect the time-series characteristics of power consumption behavior, and limits the prediction accuracy. Although existing time-series prediction methods can well mine the time-series dependence relationship of power consumption data, due to the significant time discontinuity characteristics of the historical sample set screened based on meteorological similarity, there is almost no research on combining power consumption prediction based on similar days with time-series prediction methods, and the improvement of the prediction result accuracy of daily power consumption is limited. Summary of the Invention

[0007] The purpose of the present invention is to solve the problems existing in the prior art and improve the prediction accuracy of daily power consumption.

[0008] The present invention adopts the following technical solutions to achieve the above purpose: A recursive multi-step prediction method for daily power consumption considering meteorologically similar days, comprising the following steps: Step S10: Obtain the historical daily meteorological element data with hourly resolution and the corresponding hourly electricity consumption data in the target area. Through correlation analysis, perform dimensionality reduction on the meteorological elements and screen out the key meteorological elements.

[0009] Step S20: Use the key meteorological elements to calculate the comprehensive similarity. Select n historical days as the meteorological similar days for the day to be predicted according to the comprehensive similarity.

[0010] Step S30: Extract and organize the data of the similar days and the day before their time series, form the training set data and group them.

[0011] Step S40: Train a neural network using the grouped data.

[0012] Step S50: Use the trained neural network, take the data of the first 24 hours of the day to be predicted as the input, perform recursive prediction, and obtain the prediction result of the electricity consumption on the day to be predicted.

[0013] Further, in step S10, the Pearson correlation coefficient method is used to calculate the correlation between all meteorological elements and the electricity consumption data, analyze the calculation results of the correlation of different meteorological elements, and select the km meteorological elements with the highest correlation as the key meteorological elements.

[0014] In an implementable method, in step S20, calculate the similarity between each key meteorological element between the day to be predicted and the historical days, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity; sort the historical days according to the comprehensive similarity, and select the top n historical days as the meteorological similar days, where n = 125.

[0015] In another implementable method, in step S20, calculate the similarity between each key meteorological element between the day to be predicted and the historical days, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity sim1; calculate the similarity between each key meteorological element between the day before the day to be predicted and the day before the historical days, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity sim2; sort according to the size of sim1 * 75% + sim2 * 25%, and select the top n corresponding historical days as the meteorological similar days, where n = 125.

[0016] Further, the grey relational analysis method is used to calculate the similarity.

[0017] Further, in step S30, the training set data includes hourly feature vectors, and the hourly feature vectors include the electricity consumption data corresponding to a certain hour and the data of the selected key meteorological factors.

[0018] Further, in step S30, the grouping includes: for the 48-hour feature vectors of the similar day and the day before it, first take the first 24-hour feature vectors as input data, and the electricity consumption data in the 25th-hour feature vector as the output label; then take the feature vector data from the 2nd to 25th hours as input, and the electricity consumption data in the 26th-hour feature vector as the output, and so on. A total of 24 groups of training data can be divided from the data of each similar day and the day before it.

[0019] Further, in step S50, use the trained neural network, take the data of the 24 hours before the prediction day as input, perform 24 recursive predictions, obtain the prediction results of the electricity consumption for the 24 hours of the prediction day, and sum them to obtain the data of the total daily electricity consumption of the prediction day.

[0020] The present invention also proposes a recursive multi-step prediction device for daily electricity consumption considering meteorological similar days to implement the above method.

[0021] The device includes: Meteorological data processing module: Obtain the historical daily meteorological element data and the corresponding hourly electricity consumption data of the target area with hourly resolution, perform dimensionality reduction processing on the meteorological elements through correlation analysis, and screen out key meteorological elements.

[0022] Meteorological similar day determination module: Use the key meteorological elements to calculate the comprehensive similarity, and select n historical days as the meteorological similar days of the day to be predicted according to the comprehensive similarity.

[0023] Training set data generation module: Extract and organize the data of the similar day and the day before its time series, form the training set data and group it.

[0024] Neural network training module: Use the grouped data to train the neural network.

[0025] Prediction module: Use the trained neural network, take the data of the 24 hours before the prediction day as input, perform recursive prediction, and obtain the prediction results of the electricity consumption of the prediction day.

[0026] The effective benefits of the present invention are: (1) The meteorological similar day selection algorithm not only uses historical days with meteorological conditions similar to the prediction day to construct the training sample set, so that the model can focus on the electricity consumption patterns under specific meteorological conditions, improve the applicability of the model in different scenarios, but also takes into account the differences in the influence degrees of various meteorological factors on electricity consumption, and may mask key factors due to the high correlation of non-key meteorological factors, resulting in the problem of inconsistent electricity consumption patterns despite high similarity. Therefore, a meteorological similar day selection algorithm based on comprehensive similarity is proposed, which improves the accuracy of meteorological similar day selection and provides a reliable data basis for subsequent electricity consumption prediction.

[0027] (2) By using similar day data that is not temporally continuous to train the prediction model, not only the advantages of the similar day algorithm are retained, the applicability of the model is improved, but also the sequential pattern of electricity consumption behavior can be considered, the time dependence of electricity consumption can be captured, and the prediction accuracy of daily electricity consumption is further optimized.

[0028] (3) Using the data of similar days and the previous day for training can effectively capture the short-term dynamic characteristics of the power load. And using the data of the previous day for training and prediction not only retains the same weather conditions and user behavior patterns as the prediction day, but also can accurately inherit the regular characteristics such as peak electricity consumption periods. Description of the Drawings

[0029] Figure 1 is the flow chart of the method of the present invention; Figure 2 is the Pearson correlation analysis result of the electricity consumption data and all meteorological factor data in the dataset based on this method; Figure 3 is the schematic diagram of grouping and processing meteorologically similar days and the data of the previous day in time sequence in the method of the present invention; Figure 4 is the curve graph of the prediction result of the total daily electricity consumption for 15 days of a certain substation area by applying this method; Figure 5 is the curve graph of the prediction results of the total daily electricity consumption for 15 days by applying multiple different prediction methods; Figure 6 is the summary of the prediction result accuracies of different prediction methods; Figure 7 is the schematic diagram of the composition of the device proposed by the present invention. Detailed Embodiment

[0030] The following further describes the present invention in conjunction with the drawings in the specification and the preferred embodiments, but does not limit the protection scope of the present invention thereby.

[0031] The application scenarios of the present invention are mainly for group users, such as a substation area, a park, a large or medium-sized enterprise, a residential community, etc.

[0032] Referring to Figure 1 , the present invention proposes a recursive multi-step prediction method for daily electricity consumption considering meteorologically similar days, including the following steps: Step S10: Obtain the historical daily meteorological element data and the corresponding hourly electricity consumption data with an hourly resolution in the target area, and through correlation analysis, perform dimensionality reduction processing on the meteorological elements to screen out key meteorological elements.

[0033] Step S20: Calculate the comprehensive similarity using key meteorological elements, and select n historical days as the meteorological similar days for the day to be predicted according to the comprehensive similarity.

[0034] Step S30: Extract and organize the data of the similar days and the day before their time series, form the training set data and group them.

[0035] Step S40: Train a neural network using the grouped data.

[0036] Step S50: Use the trained neural network, take the data of the first 24 hours of the prediction day as the input, and perform recursive prediction to obtain the prediction result of the electricity consumption on the prediction day.

[0037] In step S10, the target area is the area where the electricity consumption for the day to be predicted is required.

[0038] The time range of the data is generally about one year. If the amount of data is small and cannot meet the requirements, the time range can also be expanded, such as three years of data.

[0039] In this embodiment, temperature, humidity, wind speed, air pressure, and wind direction data with hourly resolution from January 1, 2011 to December 20, 2011 of the target area composed of 300 users and the electricity consumption data with hourly resolution in the corresponding time range are obtained. The correlation analysis between different meteorological element data and electricity consumption data is carried out using the data in the full time range to select the key meteorological influencing factors. Calculate the correlation degree between all meteorological element sequences and the electricity consumption sequence in the dataset using the Pearson correlation coefficient method. The calculation method is as follows: Where c is the correlation degree between variable F and variable L ; is the average value of variable F ; is the average value of variable L ; F j is the F th value of variable j ; L j is the L th value of variable j ; p is the number of data samples. In this embodiment, p = 300 * 24 = 7200.

[0040] c The absolute value range of cThe larger the absolute value, the stronger the correlation between the two variables.

[0041] In the above formula, F and L represent a certain meteorological element and electricity consumption respectively.

[0042] The calculation results of the correlation degrees between each meteorological element and electricity consumption are as follows: Please refer to Figure 2 , which shows the Pearson correlation analysis results of the electricity consumption data and all meteorological factor data in the dataset based on this method. In the coordinate axes: EC is electricity consumption, T is temperature, p is atmospheric pressure, H is humidity, D is wind direction, S is wind speed, and the color of each grid in the figure is the result of the characteristic correlation analysis. The darker the color, the more correlated the two variables are.

[0043] On this basis, analyze the calculation results of the correlation degrees of different meteorological elements, and select the km meteorological elements with the highest correlation degrees as the key meteorological influencing factors. In this embodiment, km = 3, and the three key meteorological elements are: wind speed, humidity, and temperature.

[0044] The specific implementation method of step S20 is: calculate the similarity between each key meteorological element between the day to be predicted and the historical days, and perform a weighted sum of the similarities of each key meteorological element to obtain a comprehensive similarity; sort the historical days according to the comprehensive similarity, and select the top n historical days as meteorological similar days, where n = 125.

[0045] When selecting meteorological similar days, we once considered a selection strategy based on a similarity threshold, that is, setting a similarity threshold and selecting the days in history whose meteorological conditions are more similar to the target day than this threshold as the training set. However, this method has exposed some problems in practical applications, especially when longer-term predictions are required. Specifically, each time a prediction is made, a neural network with a pre-set structure (such as parameters like the number of layers and learning rate) needs to be trained using the training set composed of the similar days of the day to be predicted, and then this network is used to predict the electricity consumption of the day. Due to the different meteorological conditions of each day to be predicted, the number of similar days selected through the threshold may vary significantly. Some days may select a large number of similar days, while some days may only select a few similar days. And the neural network model is highly sensitive to the training data, and changes in the size and data distribution of the training set may cause significant fluctuations in the prediction results of the model, thereby affecting the stability and reliability of the prediction.

[0046] In the present invention, 125 meteorologically similar days are selected as the training set. This selection is based on the marginal effect analysis of similarity calculation and the results of experimental verification. Specifically, as the number of meteorologically similar days increases, the improvement effect of the newly added days on the model prediction performance gradually weakens, showing an obvious feature of diminishing marginal effect. Through experimental verification, we found that when the size of the training set reaches 125 days, the prediction error of the model reaches the minimum value, and the model performance reaches the optimal balance point. Before this balance point, as the size of the training set increases, the prediction error of the model decreases significantly; after this point, the further reduction of the prediction error brought about by continuously increasing the size of the training set is very limited, and it may lead to a decline in model performance due to the introduction of too much noise data. Therefore, selecting a training set size of 125 days is the optimal choice through analysis and experimental verification, and the test results also prove that this selection method is feasible.

[0047] In this embodiment, the grey relational analysis method is used to calculate the similarity.

[0048] The calculation method includes: Use to represent the data vector of the th key meteorological factor of the prediction day, to represent the data vector of the th historical day and the th key meteorological factor: where represents the data of the th historical day, the th feature, and the th hour.

[0049] Use the following formula to calculate the similarity between the historical day and the prediction day: where is the resolution coefficient, with a value ranging from 0 to 1. Generally, the value is 0.5. In this embodiment takes 0.5; is the similarity between the prediction day and the i th historical day and the th meteorological element.

[0050] Perform a weighted sum of the similarities of each key meteorological element to obtain the comprehensive similarity.

[0051] The calculation method is as follows: Utilize the calculation result of the relevance of the key meteorological factors in step S10 to assign weights to each meteorological element.

[0052] Among them, for the correlation between the key meteorological factors and the electricity consumption, is the weight corresponding to the key meteorological factor.

[0053] Calculate the comprehensive similarity of the nth historical day: Among them is the comprehensive similarity.

[0054] Sort the historical days according to the comprehensive similarity, and select the top 125 historical days as meteorologically similar days.

[0055] The following table shows the calculation results of the comprehensive similarity of historical days when selecting meteorologically similar days for December 6, 2011: In the above algorithm for selecting meteorologically similar days, on the one hand, it considers the meteorological conditions similar to the prediction day, selects the key meteorological factors, and on this basis, selects the meteorologically similar days based on the comprehensive similarity, improving the accuracy of selecting meteorologically similar days and providing a reliable data basis for subsequent electricity consumption prediction.

[0056] In the method proposed by the present invention, the data of the historical day and the day before the historical day need to be used for model training. In the above method, only the historical days similar to the day to be predicted are considered.

[0057] In order to more accurately select similar days, the present invention also proposes the following embodiments: In step S20, calculate the similarity between each key meteorological element between the day to be predicted and the historical day, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity sim1; Calculate the similarity between each key meteorological element between the day before the day to be predicted and the day before the historical day, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity sim2; Weight the two comprehensive similarities, sort according to the size of sim1*75% + sim2*25%, and select the top n corresponding historical days as meteorologically similar days, where n = 125. <\(

[0058] The calculation method of the comprehensive similarity is as described above.

[0059] In step S30, the training set data includes hourly feature vectors, and the hourly feature vectors include the electricity consumption data corresponding to a certain hour and the data of the selected key meteorological factors.

[0060] Construct the hourly feature vector: The electricity consumption data corresponding to a certain hour and the data of the key meteorological factors form a vector, and this vector is the hourly feature vector. Organize the data of the similar day and the day before it into the form of 48 hourly feature vectors, and perform this operation on all the selected similar days.

[0061] Grouping: For the 48 hourly feature vectors of the similar day and the day before it, first take the first 24 hourly feature vectors as input data, and the electricity consumption data in the 25th hourly feature vector as the output label; then take the data of the 2nd to 25th hourly feature vectors as input, and the electricity consumption data in the 26th hourly feature vector as the output, and so on. The data of each similar day and the day before it can be divided into 24 groups of training data, as Figure 3 shown. Among them, a 1(i), b 1(i), e 1(i), g 1(i) represent the wind speed, humidity, temperature, and electricity consumption at the i-th hour of the day before the similar day respectively, a 2(i), b 2(i), e 2(i), g 2(i) represent the wind speed, humidity, temperature, and electricity consumption at the i-th hour of the similar day respectively.

[0062] Using the data of the similar day and the day before it for training can effectively capture the short-term dynamic characteristics of the power load.

[0063] Finally, the same processing is performed on the selected 125 similar days, and finally 3000 training sample data can be obtained.

[0064] In step S40, use the organized grouped data to train the LSTM neural network, and continuously adjust the model parameters during the training process to make the performance of the model optimal.

[0065] The input of the neural network is the key meteorological data and electricity consumption data per hour in the 24 hours before the hour to be predicted, and the output label is the electricity consumption of the hour to be predicted.

[0066] On this basis, perform step S50: Use the trained neural network, take the data of the 24 hours before the prediction day as input, and perform recursive prediction to obtain the prediction result of the electricity consumption of the prediction day.

[0067] Specifically as follows: The data of the first 24 hours of the forecast day time series are input into the trained neural network to obtain the electricity consumption data of the first hour of the forecast day. The weather forecast data corresponding to the first hour is combined with the electricity consumption data to form the hourly feature vector of the first hour. The feature vectors of the first 23 hours of the forecast day and the feature vector data of the first hour are combined as input to obtain the predicted electricity consumption data of the second hour. This process is repeated 24 times to obtain the predicted electricity consumption results for the 24 hours of the forecast day, and the total daily electricity consumption data for the forecast day is obtained by summing them up.

[0068] When obtaining the total daily electricity consumption, 24 recursive predictions were performed, that is, the total daily electricity consumption was obtained by predicting the next 24 time steps. Therefore, this prediction method is called recursive multi-step prediction.

[0069] Using electricity consumption from the day before the forecast date as model input effectively captures the short-term dynamics of power load. Because the correlation between electricity consumption data decreases rapidly with increasing time intervals (much like how memory fades over time), data closer to the forecast date (e.g., within 24 hours) not only retains the same weather conditions and user behavior patterns as the forecast date, but also accurately inherits regular patterns such as peak hours. This "nearby advantage" enables the model to capture both the regular patterns of daily electricity consumption (such as morning peaks and evening household electricity consumption) and the sudden changes it detects. For example, using yesterday's temperature to predict today is more reliable than using last week's data, as the impact of sudden events such as cold waves, heavy rains, or temporary power cuts on electricity consumption is more directly and clearly reflected in data from adjacent days. Furthermore, users are likely to use the same electrical equipment on the forecast date as on the previous day, further ensuring consistent electricity usage patterns.

[0070] When predicting mid-term daily electricity consumption, repeat steps S20-S50 fifteen times to obtain mid-term daily electricity consumption forecast results for the next 1-15 days. If there is no real data for the day before the day to be predicted, the forecast data for the previous day is used as the previous data for predicting the day to be predicted.

[0071] The prediction method proposed in this invention is used to predict the daily total electricity consumption data from December 6 to December 20, 2011 on the data set of a target area from January 1, 2011 to December 20, 2011. The curve of the prediction result is shown in the attached figure. Figure 4 shown.

[0072] For verification and comparison, the following method is used to predict the above data set: Method 2: Select similar days using the meteorological similar day selection method proposed in the present invention, but do not use the recursive multi-step prediction method proposed in the present invention. Instead, use a traditional daily total power consumption prediction method to predict the daily total power consumption of the target prediction day. This method first uses the meteorological similar day selection method proposed in the present invention; secondly, uses the data of the selected 125 meteorological similar days to train an LSTM neural network. The model input is the 24-hour average of the 3 key meteorological element data of the similar day, and the output label is the daily total power consumption data of the similar day; finally, inputs the 3 meteorological element data of the target prediction day into the trained model, and outputs the daily total power consumption data of the target prediction day.

[0073] Method 3: Use a traditional meteorological similar day selection method and a recursive multi-step prediction strategy to predict the daily total power consumption data of the target prediction day, without using the meteorological similar day selection method proposed in the present invention. The traditional meteorological similar day selection method is as follows: First, use the grey relational coefficient method to calculate the similarity between all meteorological elements, namely temperature, humidity, wind speed, air pressure, and wind direction data, of the target prediction day and each historical day; secondly, take the average of the similarity calculation results of all meteorological elements, and the obtained calculation result is the similarity between the historical day and the target prediction day; finally, select the top 125 days with the highest similarity ranking as similar days according to the similarity. Extract and organize the data of the selected 125 similar days and the data of the previous day of their time series to form the training set data, and train a long short-term memory (LSTM) neural network. The input of the neural network is the key meteorological data and power consumption data per hour of the previous 24 hours of the hour to be predicted, and the output label is the power consumption of the hour to be predicted. Finally, use the data of the previous 24 hours of the target prediction day as the input, and perform 24 recursive predictions to obtain the prediction results of the power consumption of 24 hours of the prediction day, and sum to obtain the data of the daily total power consumption of the prediction day.

[0074] Method 4: Do not use any similar day selection strategies disclosed in the present invention and also do not use a recursive multi-step prediction strategy. This method directly uses all historical days in the dataset as training samples to train a neural network. The model input is the 24-hour average of all meteorological element data of the historical day, namely temperature, humidity, wind speed, air pressure, and wind direction data, and the output label is the daily total power consumption data of the historical day; finally, input the 24h average of the 5 meteorological element data of the target prediction day into the trained model, and the output is the daily total power consumption data of the target prediction day.

[0075] The accuracy of the prediction results is measured and evaluated using three commonly used error metric indicators in prediction problems: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The smaller the values of MAE, RMSE, and MAPE, the smaller the prediction error and the higher the accuracy of the prediction results.

[0076] The prediction method proposed in the present invention has better performance compared with the other three methods. See Figure 5 、 6 。

[0077] Among them, compared with Method 2, the MAPE of the prediction method proposed in the present invention is only 3.396%, which is 4.416% lower than that of Method 2. This shows that the recursive multi-step prediction using hourly electricity consumption data proposed in the present invention can indeed improve the accuracy of the prediction results. This is because using the recursive multi-step prediction method, the neural network can learn the temporal pattern of electricity consumption behavior through the recursive multi-step strategy and capture the intra-day electricity fluctuations. At the same time, compared with Method 3, the MAPE of the prediction method proposed in the present invention is reduced by 2.395%, which shows that the method for selecting meteorologically similar days proposed in the present invention can further improve the accuracy of the prediction results. The comparison with the results of Method 4 also shows this point. This is because the weather conditions, load trends, etc. of the similar days are more similar to those of the target prediction day, and using these data to train the neural network makes the neural network more targeted.

[0078] The innovation of the present invention is to combine meteorologically similar days with recursive prediction for the prediction of daily total electricity consumption.

[0079] Meteorologically similar days are selected from historical data, and the dates of the selected meteorologically similar days are discontinuous. For example, one meteorologically similar day may be January 1st, another may be February 1st, and another may be February 5th. This phenomenon is called discontinuous in time series. However, the recursive prediction method is used for predicting continuous time series. Currently, there is no research on combining these two methods for the prediction of daily total electricity consumption. At the same time, combining these two methods retains the advantages of traditional prediction based on meteorologically similar days, that is, the model is more targeted for learning in specific situations, and also overcomes the defects of traditional prediction based on meteorologically similar days, that is, using daily total electricity consumption data for prediction cannot take into account the intra-day electricity fluctuations, and the improvement of the prediction result accuracy is limited.

[0080] Using similar data from the previous day is to combine the two methods. This method uses hourly data for recursive prediction. To use recursive prediction to predict the data of the first hour of a similar day, it is necessary to have the data of several hours before this hour.

[0081] Intraday electricity consumption fluctuation: It means that the change in electricity consumption during a day is also regular. For example, for residential users, from 8:00 to 11:00 and from 14:00 to 18:00 are working hours, and the electricity consumption of users during this period may be relatively low. However, during periods such as evenings when residents are at home, their electricity consumption will increase.

[0082] The present invention also presents an embodiment of a device for recursive multi-step prediction of daily electricity consumption considering meteorologically similar days. Refer to Figure 7 .

[0083] The device includes: Meteorological data processing module: Obtain historical daily meteorological element data and corresponding hourly electricity consumption data of the target area with hourly resolution. Through correlation analysis, perform dimensionality reduction processing on meteorological elements and screen key meteorological elements; Meteorologically similar day determination module: Use key meteorological elements to calculate the comprehensive similarity, and select n historical days as the meteorologically similar days of the day to be predicted according to the comprehensive similarity; Training set data generation module: Extract and organize the data of similar days and the previous day of their time series, form training set data and group them; Neural network training module: Use the grouped data to train the neural network; Prediction module: Use the trained neural network, take the data of the previous 24 hours of the prediction day as the input, and perform recursive prediction to obtain the prediction result of the electricity consumption of the prediction day.

Claims

1. A recursive multi-step prediction method for daily electricity consumption considering meteorological similar days, characterized in that, It includes the following steps: Step S10: Obtain the historical daily meteorological element data with an hourly resolution and the corresponding hourly electricity consumption data in the target area. Through correlation analysis, perform dimensionality reduction processing on the meteorological elements and screen out the key meteorological elements; Step S20: Use the key meteorological elements to calculate the comprehensive similarity. According to the comprehensive similarity, select n historical days as the meteorological similar days for the day to be predicted; Step S30: Extract and organize the data of the similar days and the previous day of their time series, form the training set data and group them; Step S40: Use the grouped data to train the neural network; Step S50: Use the trained neural network, take the data of the first 24 hours of the prediction day as the input, perform recursive prediction, and obtain the prediction result of the electricity consumption on the prediction day.

2. The prediction method according to claim 1, wherein In the step S10, the Pearson correlation coefficient method is used to calculate the correlation degree between all meteorological elements and the electricity consumption data, analyze the calculation results of the correlation degrees of different meteorological elements, and select the km meteorological elements with the highest correlation degree as the key meteorological elements.

3. The prediction method according to claim 1, wherein In the step S20, calculate the similarity between each key meteorological element between the day to be predicted and the historical days, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity; sort the historical days according to the comprehensive similarity, and select the top n historical days as the meteorological similar days, n = 125.

4. The prediction method according to claim 1, characterized in that In the step S20, calculate the similarity between each key meteorological element between the day to be predicted and the historical days, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity sim1, calculate the similarity between each key meteorological element between the previous day of the day to be predicted and the previous day of the historical days, and perform weighted summation on the similarities of each key meteorological element to obtain the comprehensive similarity sim2, sort according to the size of sim1 * 75% + sim2 * 25%, and select the top n corresponding historical days as the meteorological similar days, n = 125.

5. The prediction method according to claim 3 or 4, characterized in that Use the grey relational analysis method to calculate the similarity.

6. The prediction method according to claim 1, wherein In the step S30, the training set data includes hourly feature vectors, and the hourly feature vectors include the electricity consumption data corresponding to a certain hour and the data of the selected key meteorological factors.

7. The prediction method according to claim 6, characterized in that, In the step S30, the grouping includes: for the 48 hourly feature vectors of the similar day and the previous day, first take the first 24 hourly feature vectors as the input data, and the electricity consumption data in the 25th hourly feature vector as the output label; then take the data of the 2nd to 25th hourly feature vectors as the input, and the electricity consumption data in the 26th hourly feature vector as the output, and so on. The data of each similar day and the previous day can be divided into 24 groups of training data in total.

8. The prediction method according to claim 1, characterized in that In the step S50, use the trained neural network, take the data of the first 24 hours before the prediction day as the input, perform 24 recursive predictions, obtain the prediction results of the electricity consumption for 24 hours on the prediction day, and sum them to obtain the data of the total daily electricity consumption on the prediction day.

9. The prediction method according to claim 1, wherein When predicting the medium-term daily electricity consumption, repeat steps S20 - S50 fifteen times to obtain the medium-term daily electricity consumption prediction results for the next 1 - 15 days respectively.

10. A daily electricity consumption recursive multi-step prediction device considering meteorological similar days, characterized in that, The device includes: Meteorological data processing module: Obtain the historical daily meteorological element data with hourly resolution and the corresponding hourly electricity consumption data in the target area, perform dimensionality reduction processing on the meteorological elements through correlation analysis, and screen out key meteorological elements; Meteorologically similar day determination module: Use the key meteorological elements to calculate the comprehensive similarity, and select n historical days as the meteorologically similar days of the day to be predicted according to the comprehensive similarity; Training set data generation module: Extract and organize the data of the similar days and the day before their time series, form the training set data and group them; Neural network training module: Use the grouped data to train the neural network; Prediction module: Use the trained neural network, take the data of the first 24 hours of the prediction day as the input, perform recursive prediction, and obtain the prediction result of the electricity consumption on the prediction day.

Images