Power load prediction system

Abstract

The invention belongs to the technical field of electric power prediction, and particularly relates to an electric power load prediction system which obtains a load parameter sequence through an acquisition module and extracts correlation coupling characteristics representing influence degrees of seasons, meteorology and electricity price factors; using a first prediction module to obtain a short-term load prediction result based on the sliding time window and the short-term prediction sub-model, and using a second prediction module to obtain a long-term load prediction result based on the longer time window and the long-term prediction sub-model; a coupling loss function is innovatively constructed through an online coupling module for cooperative training, and deep correlation analysis of fluctuation mode types corresponding to temperature, electricity price and load fluctuation sequences is realized by adopting fluctuation mode clustering, multi-dimensional feature extraction and load mode coupling weight matrix technologies. The technical problem of mutual isolation of long-term and short-term prediction is effectively solved, and the accuracy and adaptive capacity of load prediction are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of power forecasting technology, and in particular relates to a power load forecasting system. Background Technology

[0002] Electricity load forecasting is a core component of the safe and economical operation of the power grid. However, existing technologies currently operate long-term and short-term forecasting systems independently, presenting fundamental technical bottlenecks. On the one hand, short-term forecasting, due to its limited time horizon, cannot integrate medium- and long-term seasonal and trend-based electricity consumption patterns, leading to systematic deviations during critical periods such as seasonal transitions. On the other hand, long-term forecasting, constrained by the lag and macro-level nature of historical data, struggles to detect real-time weather changes and adjustments to electricity consumption strategies, resulting in inherent large deviations. More critically, the two systems lack an effective coordination mechanism. The real-time accuracy of short-term forecasting cannot be used to correct macro-level trends in long-term forecasting, while the strategic framework of long-term forecasting cannot dynamically constrain micro-level adjustments in short-term forecasting. This technical deficiency of mutual isolation and inability to form a closed-loop optimization severely restricts the fundamental improvement of load forecasting accuracy. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a power load forecasting system. This system acquires load parameter sequences through an acquisition module and extracts correlation and coupling features characterizing the influence of seasonal, meteorological, and electricity price factors. A first forecasting module obtains short-term load forecasting results based on a sliding time window and a short-term forecasting sub-model, while a second forecasting module obtains long-term load forecasting results based on a longer time window and a long-term forecasting sub-model. Innovatively, an online coupling module constructs a coupling loss function for collaborative training. Furthermore, it employs fluctuation pattern clustering, multi-dimensional feature extraction, and load pattern coupling weight matrix techniques to achieve deep correlation analysis of the fluctuation pattern types corresponding to temperature, electricity price, and load fluctuation sequences. This effectively solves the technical problem of the isolation between short-term and long-term forecasts, significantly improving the accuracy and adaptability of load forecasting.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A power load forecasting system, comprising:

[0006] The data acquisition module is used to acquire load parameter sequences according to a preset sampling period length and combine them with a time series algorithm to obtain correlation coupling characteristics; the correlation coupling characteristics are used to characterize the degree of influence of seasonal factors, meteorological factors and electricity price fluctuation factors on load changes.

[0007] The first prediction module is used to obtain the first load prediction result and the first prediction error, which characterize the short-term prediction, based on the load parameter sequence, a preset first sliding time window, and related coupling characteristics, through a preset short-term prediction sub-model.

[0008] The second prediction module is used to obtain a second load prediction result and a second prediction error that characterize long-term prediction by combining the load parameter sequence with a second preset time window and the associated coupling characteristics through a preset long-term prediction sub-model.

[0009] The online coupling module performs online coupling training and prediction based on the short-term prediction sub-model, the long-term prediction sub-model, the integrated framework, and the long-term and short-term coupling loss functions constructed by combining the first prediction error, the second prediction error, and the prediction coupling errors corresponding to the first load prediction result and the second load prediction result within the first sliding time window, to obtain the corresponding first load prediction result or second load prediction result.

[0010] Specifically, the preset sampling period length is greater than the second preset time window length; the second preset time window length is greater than the first sliding time window length; the process of acquiring the associated coupling feature includes:

[0011] Collect temperature and electricity price data for the corresponding time periods of the load demand time series, and perform standardized preprocessing.

[0012] Based on the processed load demand time series combined with the differential algorithm, a load differential series is obtained;

[0013] Based on the load difference sequence combined with the dynamic time warping clustering algorithm and the preset fluctuation pattern types, fluctuation pattern similarity identification and clustering are performed to obtain fluctuation pattern clustering results; the fluctuation pattern types include stationary fluctuation patterns with a difference result of 0, periodic fluctuation patterns, and transient fluctuation patterns.

[0014] Based on the load differential sequence under the steady fluctuation mode, the length of each steady fluctuation mode interval and the steadyness score are extracted; the steadyness score is equal to 1 minus the number of abnormal fluctuation points sampled within the steady fluctuation mode interval divided by the total number of sampling points.

[0015] Specifically, the process of obtaining the associated coupling features also includes:

[0016] Based on the load differential sequence under each cycle fluctuation mode and combined with Fourier transform, a frequency domain feature set under each cycle fluctuation mode is extracted; the frequency domain feature set includes the main cycle length and cycle intensity; the process of obtaining the cycle intensity includes: performing Fourier transform on the load differential sequence to obtain a frequency domain representation, identifying the frequency component with the largest amplitude as the main frequency and obtaining its amplitude, calculating the sum of squares of the amplitudes of all frequency components as the total signal energy, and taking the ratio of the amplitude of the main frequency to the total signal energy as the cycle intensity;

[0017] Simultaneously, based on the load differential sequence under each cycle fluctuation mode and combined with statistical analysis algorithms, the time-domain feature set corresponding to each cycle fluctuation mode interval is obtained; the time-domain feature set includes the average rise or fall rate, peak or trough duration, fluctuation amplitude, and phase consistency characteristics.

[0018] Based on the load differential sequence of each transient fluctuation mode and combined with statistical algorithms, a transient time-domain feature set is obtained; the transient time-domain feature set includes the mutation amplitude, mutation duration, mutation direction and mutation steepness.

[0019] Specifically, the process of obtaining the associated coupling features also includes:

[0020] Based on the stability score of each steady fluctuation mode in continuous time, the frequency domain feature set and time domain feature set of each periodic fluctuation mode, and the transient time domain feature set of each transient fluctuation mode, combined with the temperature sequence and corresponding temperature difference sequence, electricity price sequence and electricity price fluctuation sequence under the same continuous time, the contribution of temperature and electricity price variables to the fluctuation in each fluctuation mode interval is obtained by factor analysis algorithm.

[0021] Based on the load differential sequence over continuous time, with the turning point of each fluctuation mode as the center of a preset intercept time window, and the standard deviation of the load differential sequences corresponding to both sides of the turning point of the fluctuation mode as the length of the intercept time window, the cross-modal load differential sequence of the preset intercept time window length is intercepted; the standard deviation length is specifically: ,in The length of the data truncation within the i-th fluctuation mode interval of the time window. To determine the data truncation length within the (i+1)th fluctuation pattern interval of the time window, and The standard deviations are, in order, the magnitudes of the standard deviations of the i-th fluctuation pattern interval and the (i+1)-th fluctuation pattern interval; and These are the interval lengths of the i-th fluctuation pattern interval and the (i+1)-th fluctuation pattern interval, respectively.

[0022] Specifically, the process of obtaining the associated coupling features also includes:

[0023] Based on the cross-modal load differential sequence combined with the temperature differential sequence and electricity price fluctuation sequence in the corresponding time period, the first causal correlation degree of temperature fluctuation on load fluctuation and the second causal correlation degree of electricity price fluctuation on load fluctuation in the corresponding time period are obtained through causal correlation analysis algorithm.

[0024] The contribution of temperature and electricity price variables to the fluctuation within each fluctuation mode interval is standardized by combining the first causal correlation between temperature fluctuation and load fluctuation and the second causal correlation between electricity price fluctuation and load fluctuation in the corresponding time period, and then arranged according to the timestamp of the corresponding fluctuation mode interval to obtain the load mode coupling weight matrix.

[0025] Based on the load pattern coupling weight matrix, combined with the temperature fluctuation sequence, the electricity price fluctuation sequence, and the fluctuation pattern type corresponding to the load fluctuation sequence, a convolutional attention network is used for convolutional coupling to obtain the correlation coupling features that characterize the degree of influence of temperature and electricity price on the fluctuation pattern type corresponding to the load fluctuation sequence.

[0026] Specifically, the construction process of the short-term forecast sub-model includes:

[0027] Obtain short-term load forecast demand and load differential information, temperature and electricity price information for a preset short-term time length, and extract the corresponding demand forecast time length, forecast timestamp information, and current temperature fluctuation sequence and electricity price fluctuation sequence.

[0028] Based on the demand forecast time length, forecast timestamp information, load differential information, temperature fluctuation sequence and electricity price fluctuation sequence combined with the aforementioned correlation and coupling features, a pattern recognition algorithm is used to identify the fluctuation patterns contained within the demand forecast time length, the corresponding pattern interval length, the fluctuation pattern turning point, and the timestamp position of each current pattern interval length in the load pattern coupling weight matrix.

[0029] The fluctuation contribution of each identified fluctuation pattern at the timestamp position in the load pattern coupling weight matrix is ​​extracted as the training weight within the fluctuation pattern. At the same time, the first causal correlation degree and the second causal correlation degree corresponding to the intercept time window at the turning point of each fluctuation pattern are extracted as cross-mode training weights to construct a short-term training weight vector.

[0030] Specifically, the construction process of the short-term forecast sub-model also includes:

[0031] The first sliding time window for the corresponding fluctuation pattern interval is used as the ratio of the length of each identified fluctuation pattern interval to the standard deviation of the corresponding interval. The training input sequence set is obtained based on the load differential information combined with the length of the first sliding time window.

[0032] Based on the training input sequence set, short-term training weight vector, and gradient boosting tree, training is performed within the fluctuation mode interval and across modes to obtain the trained short-term prediction sub-model, and the first load prediction result and the first prediction error corresponding to the demand prediction time length are output.

[0033] Specifically, the construction process of the long-term prediction sub-model includes:

[0034] Obtain long-term forecast demand, load differential information, temperature and electricity price information for at least one year, and extract the corresponding forecast timestamp information, long-term forecast duration, current load differential information, temperature fluctuation sequence and electricity price fluctuation sequence;

[0035] Based on the predicted timestamp information and the current temperature fluctuation sequence and electricity price fluctuation sequence, combined with the correlation coupling feature and vector error correction model, the system is trained in the manner of training within the fluctuation mode interval and cross-mode training, and outputs the second load prediction result and the second prediction error.

[0036] Specifically, the construction process of the long-term prediction sub-model also includes:

[0037] While making long-term forecasts, synchronous fluctuation patterns are identified based on the second load forecast results obtained during the forecasting process. The trained short-term forecast sub-model is then called to perform different mode load forecasts for each preset short-term time length under the long-term forecast time length, so as to obtain the synchronous first load forecast results under the second load forecast results.

[0038] Based on the synchronous first load forecast result and the second load forecast result under the corresponding preset short-term time length, the forecast coupling error corresponding to the long-term forecast is obtained.

[0039] When the prediction coupling error is 0 and the second prediction error meets the preset long-term error threshold, the trained long-term prediction sub-model is obtained.

[0040] Specifically, in cross-mode training, the left and right boundaries of each intercepted time window are used as the input starting points for two different fluctuation modes. Combining the load differential information and cross-mode training weights corresponding to the two sides of the fluctuation mode inflection point, synchronous sliding input training is performed from the left boundary to the fluctuation mode inflection point and from the right boundary to the fluctuation mode inflection point, respectively, according to the input length of the first sliding time window of the corresponding fluctuation mode interval.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention addresses the shortcomings of existing technologies by innovatively integrating short-term and long-term load forecasting systems and introducing an online coupling module to construct a collaborative training mechanism. This effectively solves the fundamental bottleneck of existing technologies where short-term and long-term forecasts are isolated and unable to form a closed-loop optimization. The system extracts the correlation and coupling characteristics of factors such as seasons, weather, and electricity prices through a data acquisition module, and uses fluctuation pattern clustering and multi-dimensional analysis techniques to deeply explore the intrinsic relationship between temperature, electricity prices, and load fluctuations. The short-term forecasting module captures micro-changes based on a sliding time window and real-time data, while the long-term forecasting module focuses on macro-trends. The online coupling module achieves dynamic interaction between the two through a coupling loss function, enabling the real-time accuracy of short-term forecasts to correct macro-level deviations in long-term forecasts, while the strategic framework of long-term forecasts constrains the micro-level adjustments of short-term forecasts. This collaborative mechanism significantly improves the overall accuracy and adaptability of load forecasting, especially during critical periods such as seasonal transitions, reducing systemic deviations, enhancing the safety and economy of power grid operation, and providing more reliable data support for power system planning and dispatch. Attached Figure Description

[0043] Figure 1 This is a block diagram of a power load forecasting system according to Embodiment 1 of the present invention;

[0044] Figure 2 This is a schematic diagram of capturing a time window in Embodiment 1 of the present invention. Detailed Implementation

[0045] Example 1

[0046] Please see Figure 1 The present invention provides an embodiment of a power load forecasting system, comprising the following steps:

[0047] The data acquisition module is used to acquire load parameter sequences according to a preset sampling period length and combine them with a time series algorithm to obtain correlation coupling characteristics; the correlation coupling characteristics are used to characterize the degree of influence of seasonal factors, meteorological factors and electricity price fluctuation factors on load changes.

[0048] The first prediction module is used to obtain the first load prediction result and the first prediction error, which characterize the short-term prediction, based on the load parameter sequence, a preset first sliding time window, and related coupling characteristics, through a preset short-term prediction sub-model.

[0049] The second prediction module is used to obtain a second load prediction result and a second prediction error that characterize long-term prediction by combining the load parameter sequence with a second preset time window and the associated coupling characteristics through a preset long-term prediction sub-model.

[0050] The online coupling module performs online coupling training and prediction based on the short-term prediction sub-model, the long-term prediction sub-model, the integrated framework, and the long-term and short-term coupling loss functions constructed by combining the first prediction error, the second prediction error, and the prediction coupling errors corresponding to the first load prediction result and the second load prediction result within the first sliding time window, to obtain the corresponding first load prediction result or second load prediction result.

[0051] It should be further noted that, in this embodiment, the preset sampling period length is greater than the second preset time window length; the second preset time window length is greater than the first sliding time window length; it should also be noted that, in this embodiment, the process of obtaining the associated coupling features includes:

[0052] Collect temperature and electricity price data for the corresponding time periods of the load demand time series, and perform standardized preprocessing.

[0053] Based on the processed load demand time series combined with the differential algorithm, a load differential series is obtained;

[0054] Based on the load difference sequence combined with the dynamic time warping clustering algorithm and the preset fluctuation pattern types, fluctuation pattern similarity identification and clustering are performed to obtain fluctuation pattern clustering results; the fluctuation pattern types include stationary fluctuation patterns with a difference result of 0, periodic fluctuation patterns, and transient fluctuation patterns.

[0055] Based on the load differential sequence under the steady fluctuation mode, the length of each steady fluctuation mode interval and the steadyness score are extracted; the steadyness score is equal to 1 minus the number of abnormal fluctuation points sampled within the steady fluctuation mode interval divided by the total number of sampling points.

[0056] Based on the load differential sequence under each cycle fluctuation mode and combined with Fourier transform, a frequency domain feature set under each cycle fluctuation mode is extracted; the frequency domain feature set includes the main cycle length and cycle intensity; the process of obtaining the cycle intensity includes: performing Fourier transform on the load differential sequence to obtain a frequency domain representation, identifying the frequency component with the largest amplitude as the main frequency and obtaining its amplitude, calculating the sum of squares of the amplitudes of all frequency components as the total signal energy, and taking the ratio of the amplitude of the main frequency to the total signal energy as the cycle intensity;

[0057] Simultaneously, based on the load differential sequence under each cycle fluctuation mode and combined with statistical analysis algorithms, the time-domain feature set corresponding to each cycle fluctuation mode interval is obtained; the time-domain feature set includes the average rise or fall rate, peak or trough duration, fluctuation amplitude, and phase consistency characteristics.

[0058] Based on the load differential sequence of each transient fluctuation mode and combined with statistical algorithms, a transient time-domain feature set is obtained; the transient time-domain feature set includes the mutation amplitude, mutation duration, mutation direction and mutation steepness.

[0059] It should be further explained that the implementation process of wave pattern similarity identification and clustering in this embodiment includes:

[0060] Step 1: Based on the load parameter sequence containing load values ​​at multiple consecutive sampling times collected according to a preset sampling period, the load difference sequence is obtained by calculating the difference between the load value at the next time moment and the load value at the previous time moment. Then, the difference sequence is standardized to eliminate the difference in data magnitude and obtain the standardized load difference sequence.

[0061] Step 2: Based on the standardized load differential sequence output in Step 1, it has been pre-divided into multiple continuous subsequences with fixed sampling points according to a preset time window, and the feature set of the subsequences is extracted. Each subsequence in this set corresponds to three sets of features: stationary correlation features, periodic correlation features and time domain features, including main period length, period intensity, mean rate, etc.; transient correlation features, including abrupt change amplitude, duration, etc.

[0062] Step 3: Based on the standardized load differential sequence from Step 1 and the feature set of the subsequence from Step 2, for any two subsequences, first construct a distance matrix of the absolute difference of the sampling points at corresponding positions, then calculate the difference degree of the feature set, i.e. the sum of the feature differences. Solve the optimal matching path distance by dynamic programming combined with the feature difference degree correction, and arrange the pairwise distances of all subsequences according to their corresponding positions to form a dynamic time-warped distance matrix.

[0063] Step 4: Based on the dynamic time-warped distance matrix in Step 3, preset 3 cluster numbers (corresponding to three types of fluctuation patterns), use the K-means clustering algorithm, with the dynamic time-warped distance as the similarity measure, and through the iterative process of randomly initializing cluster centers, assigning subsequences to the nearest cluster, and updating cluster centers, until the cluster centers are stable, obtain the preliminary clustering results containing the cluster labels of each subsequence.

[0064] Step 5: Based on the set of subsequences of a certain cluster in the preliminary clustering results of Step 4 and the stationary correlation characteristics of the corresponding cluster in Step 2, first count the number of sampling points with a difference result of 0 in the original load difference sequence of each subsequence, calculate the ratio of this number to the total number of sampling points of the subsequence, verify that the stationarity score of all subsequences is not less than 0.9, and the number of abnormal fluctuation points with a difference result of non-zero in each subsequence does not exceed 2. When the above conditions are met, the cluster is determined to be a stationary fluctuation mode, and the corresponding sequence and the time interval identifier defined by the first and last sampling times of the subsequence are obtained.

[0065] Step 6: Based on the subsequence set of another cluster in the preliminary clustering results of Step 4 and the periodic correlation features and time-domain features of the corresponding cluster in Step 2, first verify that the periodic intensity of all subsequences is greater than or equal to 0.6, and the length of the main period falls within the preset typical period range; then check the time-domain features: the coefficient of variation of the mean of the rising or falling rate, i.e., the ratio of the standard deviation to the mean, is less than or equal to 0.15; the duration of the peak or trough varies less than or equal to 5% of the length of the main period in each subsequence; the ratio of the maximum to the minimum fluctuation amplitude does not exceed 1.2; and the phase consistency is achieved by the average phase difference of three consecutive periods not exceeding 3 degrees. When all conditions are met, the cluster is determined to be a periodic fluctuation mode, and the corresponding sequence and the time interval identifier defined by the first and last sampling times of the subsequence are obtained.

[0066] Step 7: Based on the set of subsequences of the remaining cluster in the preliminary clustering results of Step 4 and the transient correlation features corresponding to this cluster in Step 2, first verify that the absolute value of the difference at the mutation point of all subsequences is not less than 3 times the global standard deviation of the load difference sequence, and the rate of change of the slope of the difference sequence before and after the mutation point, i.e., the ratio of the slope after the mutation to the slope before the mutation, exceeds 50%; then count the duration of the mutation: the number of sampling periods from the start of the mutation point to the first time the absolute value of the difference falls back to within the global standard deviation, verify that this time does not exceed 3 sampling periods, and there is no secondary mutation during the fallback process, i.e., the absolute value of the difference does not exceed 2 times the global standard deviation again; at the same time, confirm that the mutation direction (positive / negative) is consistent within the subsequence, i.e., there is no alternating positive and negative mutation. When all conditions are met, determine that the cluster is a transient fluctuation mode, and obtain the corresponding sequence and the time interval identifier defined by the first and last sampling times of the subsequence.

[0067] Based on the stability score of each steady fluctuation mode in continuous time, the frequency domain feature set and time domain feature set of each periodic fluctuation mode, and the transient time domain feature set of each transient fluctuation mode, combined with the temperature sequence and corresponding temperature difference sequence, electricity price sequence and electricity price fluctuation sequence under the same continuous time, the contribution of temperature and electricity price variables to the fluctuation in each fluctuation mode interval is obtained by factor analysis algorithm.

[0068] It should be further explained that the analysis process of the factor analysis algorithm in this embodiment includes:

[0069] Step 1: Based on the time interval identifiers corresponding to the identified stationary fluctuation pattern, periodic fluctuation pattern, and transient fluctuation pattern, divide the load difference sequence, temperature sequence, temperature difference sequence, electricity price sequence, and electricity price fluctuation sequence within a continuous time period into intervals to obtain the subset data set corresponding to each fluctuation pattern interval.

[0070] Step 2: For each fluctuation pattern interval, integrate all features and external variables to form an interval variable matrix. The variable matrix for the stationary fluctuation pattern interval includes the stationarity score, interval length, mean and standard deviation of the temperature sequence segment, mean and standard deviation of the temperature difference sequence segment, mean and standard deviation of the electricity price sequence segment, and mean and standard deviation of the electricity price fluctuation sequence segment. The variable matrix for the periodic fluctuation pattern interval includes the main period length, period intensity, mean rise rate, mean fall rate, peak duration, trough duration, fluctuation amplitude, phase consistency, and Fourier transform of the temperature sequence segment. The Fourier transform extracts the principal period and period intensity, the mean rate and fluctuation amplitude of the temperature difference sequence segment, the Fourier transform extracts the principal period and period intensity of the electricity price sequence segment, and the mean rate and fluctuation amplitude of the electricity price fluctuation sequence segment; the variable matrix of the transient fluctuation mode interval includes the mutation amplitude, mutation duration, mutation direction (positive direction is 1, negative direction is -1), mutation steepness, the maximum temperature difference and temperature difference duration of the temperature sequence segment, the mutation amplitude and steepness of the temperature difference sequence segment, the maximum price difference and price difference duration of the electricity price sequence segment, and the mutation amplitude and steepness of the electricity price fluctuation sequence segment.

[0071] Step 3: Standardize the variable matrix for each fluctuation mode interval by mapping the value of each variable in the matrix to the 0-1 interval. This is achieved by subtracting the minimum value of the variable and then dividing by the difference between the maximum and minimum values ​​of the variable. This eliminates the influence of differences in the dimensions and magnitudes of different variables, resulting in a standardized interval variable matrix.

[0072] Step 4: Perform factor analysis on the standardized interval variable matrix, calculate the correlation coefficient matrix between variables, and determine the common factors with eigenvalues ​​greater than 1. The number of common factors is determined based on the cumulative contribution rate of eigenvalues ​​being no less than 85%. Calculate the loading values ​​of each variable on the common factors, i.e., the correlation coefficients between the variables and the common factors. Among them, the loading values ​​of temperature-related variables, i.e., the characteristics of temperature series and difference series, on the common factors are integrated into temperature factor loadings, and the loading values ​​of electricity price-related variables, i.e., the characteristics of electricity price series and fluctuation series, on the common factors are integrated into electricity price factor loadings.

[0073] Step 5: Calculate the contribution of temperature and electricity price to the volatility of each volatility pattern interval based on factor loadings. For the stable volatility pattern interval, multiply the temperature factor loading by the stability score of that interval, divide by the sum of the temperature factor loading and the electricity price factor loading to obtain the temperature's contribution to the volatility of that interval. Then, multiply the electricity price factor loading by the stability score and divide by the sum to obtain the electricity price's contribution. For the periodic volatility pattern interval, using period intensity as the weight, multiply the temperature factor loading by the period intensity and divide by the sum of the products of the total load and the period intensity to obtain the temperature's contribution. Similarly, calculate the electricity price's contribution. For the transient volatility pattern interval, using the abrupt change amplitude as the weight, multiply the temperature factor loading by the abrupt change amplitude and divide by the sum of the products of the total load and the abrupt change amplitude to obtain the temperature's contribution. Similarly, calculate the electricity price's contribution. Finally, output the temperature volatility contribution and electricity price volatility contribution corresponding to each volatility pattern interval.

[0074] Based on the load differential sequence over continuous time, with the turning point of each fluctuation mode as the center of a preset intercept time window, and the standard deviation of the load differential sequences corresponding to both sides of the turning point of the fluctuation mode as the length of the intercept time window, the cross-modal load differential sequence of the preset intercept time window length is intercepted; the standard deviation length is specifically: ,in The length of the data truncation within the i-th fluctuation mode interval of the time window. To determine the data truncation length within the (i+1)th fluctuation pattern interval of the time window, and The standard deviations are, in order, the magnitudes of the standard deviations of the i-th fluctuation pattern interval and the (i+1)-th fluctuation pattern interval; and These represent the interval lengths of the i-th fluctuation pattern interval and the (i+1)-th fluctuation pattern interval, respectively. Please refer to [link / reference]. Figure 2 Where tq and ts are the left and right boundary time points of the i-th fluctuation mode interval, ts and tq are the left and right boundaries of the adjacent (i+1)-th fluctuation interval, ts is the fluctuation mode transition time point of these two fluctuation mode intervals, and ts1 and ts2 are the lengths of the preset interception time windows corresponding to the fluctuation mode transition time points. In this embodiment, the interception window is dynamically determined by the standard deviation length of the mode interval length weighted by the fluctuation mode transition time point as the center. This can accurately adapt to the interval differences of adjacent fluctuation modes, avoid the omission or redundant collection of cross-modal load features by the fixed window, and thus accurately obtain the cross-modal load differential sequence, effectively improving the feature capture accuracy and prediction accuracy of power load forecasting in mode switching scenarios.

[0075] Based on the cross-modal load differential sequence combined with the temperature differential sequence and electricity price fluctuation sequence in the corresponding time period, the first causal correlation degree of temperature fluctuation on load fluctuation and the second causal correlation degree of electricity price fluctuation on load fluctuation in the corresponding time period are obtained through causal correlation analysis algorithm.

[0076] It should be further explained that the analysis process of the causal association analysis algorithm in this embodiment includes:

[0077] Step 1: Based on the extracted cross-modal load differential sequence, and the corresponding temperature differential sequence and electricity price fluctuation sequence for the time period, classify the data according to the fluctuation mode type to obtain the cross-modal load differential sequence subsequence, temperature difference molecule sequence, and electricity price fluctuation subsequence corresponding to the stationary fluctuation mode; the cross-modal load differential sequence subsequence, temperature difference molecule sequence, and electricity price fluctuation subsequence corresponding to the periodic fluctuation mode; and the cross-modal load differential sequence subsequence, temperature difference molecule sequence, and electricity price fluctuation subsequence corresponding to the transient fluctuation mode.

[0078] Step 2: Extract preliminary association features by pattern, specifically including:

[0079] 2.1 For the cross-modal load difference subsequence, temperature difference numerator sequence, and electricity price fluctuation subsequence corresponding to the stationary fluctuation mode, calculate the correlation coefficient between load difference and temperature difference as the preliminary correlation between temperature and load; calculate the correlation coefficient between load difference and electricity price fluctuation as the preliminary correlation between electricity price and load.

[0080] 2.2 For the cross-modal load difference sequence subsequence, temperature difference numerator sequence, and electricity price fluctuation subsequence corresponding to the periodic fluctuation mode, extract the principal period length and period intensity of the load difference, temperature difference, and electricity price fluctuation; calculate the similarity between load periodic characteristics and temperature periodic characteristics (such as the weighted sum of the ratio of the period length difference to the mean period length and the ratio of the period intensity difference to the mean period intensity) as the preliminary correlation between temperature and load; similarly calculate the similarity between load periodic characteristics and electricity price periodic characteristics as the preliminary correlation between electricity price and load.

[0081] 2.3 For the cross-modal load difference sequence subsequence, temperature difference sequence, and electricity price fluctuation subsequence corresponding to the transient fluctuation mode, extract the abrupt change amplitude, abrupt change duration, and abrupt change steepness of the load difference, temperature difference, and electricity price fluctuation; calculate the matching degree between load transient characteristics and temperature transient characteristics (e.g., the weighted sum of the ratios of the difference in abrupt change amplitude to the mean of the abrupt change amplitude, the difference in abrupt change duration to the mean of the abrupt change duration, and the difference in abrupt change steepness to the mean of the abrupt change steepness), as a preliminary correlation between temperature and load; similarly, calculate the matching degree between load transient characteristics and electricity price transient characteristics, as a preliminary correlation between electricity price and load.

[0082] Step 3: Granger causality test to determine the degree of causal association, specifically including:

[0083] 3.1 Perform Granger causality tests on the load difference series and temperature difference series, and the load difference series and electricity price fluctuation series under the stationary fluctuation mode, respectively, to determine whether temperature fluctuation is a Granger cause of load fluctuation and whether electricity price fluctuation is a Granger cause of load fluctuation, and obtain the first causal correlation degree of temperature on load and the second causal correlation degree of electricity price on load under the stationary fluctuation mode.

[0084] 3.2 Perform Granger causality tests on the load differential periodic characteristic sequence under the periodic fluctuation mode, i.e. the sequence composed of the main period length and period intensity, the temperature differential periodic characteristic sequence, the load differential periodic characteristic sequence and the electricity price fluctuation periodic characteristic sequence, respectively, to obtain the first causal correlation degree of temperature on load and the second causal correlation degree of electricity price on load under the periodic fluctuation mode.

[0085] 3.3. Granger causality tests were performed on the load differential transient characteristic sequence under the transient fluctuation mode, including the sequence consisting of the abrupt change amplitude, abrupt change duration, and abrupt change steepness, and the temperature differential transient characteristic sequence, the load differential transient characteristic sequence, and the electricity price fluctuation transient characteristic sequence, respectively, to obtain the first causal correlation degree of temperature on load and the second causal correlation degree of electricity price on load under the transient fluctuation mode.

[0086] Step 4: Calculate the proportion of the cross-modal data subsequence length corresponding to the three fluctuation modes (stationary, periodic, and transient) to the total cross-modal data sequence length; multiply the first causal correlation degree under the stationary mode by its proportion, the first causal correlation degree under the periodic mode by its proportion, and the first causal correlation degree under the transient mode by its proportion, and sum them to obtain the final first causal correlation degree of temperature fluctuation on load fluctuation in the corresponding time period; similarly, multiply the second causal correlation degree under the three modes by their corresponding proportions, and sum them to obtain the final second causal correlation degree of electricity price fluctuation on load fluctuation in the corresponding time period, and output the first causal correlation degree and the second causal correlation degree.

[0087] The contribution of temperature and electricity price variables to the fluctuation within each fluctuation mode interval is standardized by combining the first causal correlation between temperature fluctuation and load fluctuation and the second causal correlation between electricity price fluctuation and load fluctuation in the corresponding time period, and then arranged according to the timestamp of the corresponding fluctuation mode interval to obtain the load mode coupling weight matrix.

[0088] Based on the load pattern coupling weight matrix, combined with the temperature fluctuation sequence, the electricity price fluctuation sequence, and the fluctuation pattern type corresponding to the load fluctuation sequence, a convolutional attention network is used for convolutional coupling to obtain the correlation coupling features that characterize the degree of influence of temperature and electricity price on the fluctuation pattern type corresponding to the load fluctuation sequence.

[0089] It should be further explained that the convolutional coupling process of the convolutional attention network in this embodiment includes:

[0090] Step 1: Based on the obtained load pattern coupling weight matrix, temperature fluctuation sequence, electricity price fluctuation sequence, and the fluctuation pattern type corresponding to the load fluctuation sequence, align the data instantaneously according to the timestamp; encode the fluctuation pattern type corresponding to the load fluctuation sequence, denoting the stable mode as 1, the periodic mode as 2, and the transient mode as 3, forming a pattern coding sequence; concatenate the temperature fluctuation contribution and electricity price fluctuation contribution in the load pattern coupling weight matrix with the temperature fluctuation sequence value, electricity price fluctuation sequence value, and pattern coding sequence value of the corresponding timestamp according to the feature dimensions to construct a multimodal input tensor. Each timestamp corresponds to a feature vector, containing five feature dimensions: temperature contribution, electricity price contribution, temperature fluctuation value, electricity price fluctuation value, and pattern coding.

[0091] Step 2: Extract features from the multimodal input tensor using a multi-branch convolutional structure, specifically including:

[0092] Stationary mode branch: Select feature vectors with mode encoding of 1 from the input tensor to form stationary mode sub-tensors; use a one-dimensional convolution kernel with a kernel size of 3 and a stride of 1 to perform convolution operation on the stationary mode sub-tensors, extract local features related to temperature, electricity price and load under stationary mode, and output stationary mode convolution feature sequence.

[0093] Periodic pattern branch: Select feature vectors with pattern encoding of 2 from the input tensor to form a periodic pattern sub-tensor; use a one-dimensional convolution kernel with a kernel size of 5 and a stride of 1 to perform a convolution operation on the periodic pattern sub-tensor to extract local features related to temperature, electricity price and load under the periodic pattern, and output the periodic pattern convolution feature sequence.

[0094] Transient mode branch: Select feature vectors with mode encoding of 3 from the input tensor to form a transient mode sub-tensor; use a one-dimensional convolution kernel with a kernel size of 2 and a stride of 1 to perform a convolution operation on the transient mode sub-tensor to extract local features related to temperature, electricity price and load under transient mode, and output the transient mode convolution feature sequence.

[0095] Step 3: Based on the contribution of temperature fluctuation and electricity price fluctuation in the load pattern coupling weight matrix, calculate the attention weights of the feature vectors for each time stamp:

[0096] For the feature vector of each timestamp, the contribution of temperature fluctuation and the contribution of electricity price fluctuation are normalized to obtain the temperature attention weight and the electricity price attention weight.

[0097] For the stationary mode convolutional feature sequence, the temperature-related features and electricity-related features in the convolutional features are weighted by the temperature attention weight and electricity price attention weight corresponding to the timestamp, respectively. Similarly, the corresponding weights are applied to the periodic mode and transient mode convolutional feature sequences to obtain the weighted stationary, periodic and transient mode feature sequences.

[0098] Step 4: Concatenate the weighted stationary, periodic, and transient mode feature sequences in time stamp order to form a fused feature sequence; use a fully connected layer to compress the dimensionality of the fused feature sequence, mapping the feature dimension to 1, to obtain the correlation coupling feature at each time stamp that characterizes the degree of influence of temperature and electricity price on the fluctuation mode type corresponding to the load fluctuation sequence, and output the correlation coupling feature sequence.

[0099] It should be further explained that the construction process of the short-term prediction sub-model in this embodiment includes:

[0100] Acquire short-term load forecast demand and load differential information, temperature and electricity price information for a preset short-term time length, and extract the corresponding demand forecast time length, forecast timestamp information, and current temperature fluctuation sequence and electricity price fluctuation sequence; the short-term load forecast demand is the short-term time length to be forecasted, which is specifically set by those skilled in the art based on the forecast demand mechanical energy.

[0101] Based on the demand forecast time length, forecast timestamp information, load differential information, temperature fluctuation sequence and electricity price fluctuation sequence combined with the aforementioned correlation and coupling features, a pattern recognition algorithm is used to identify the fluctuation patterns contained within the demand forecast time length, the corresponding pattern interval length, the fluctuation pattern turning point, and the timestamp position of each current pattern interval length in the load pattern coupling weight matrix.

[0102] The fluctuation contribution of each identified fluctuation pattern at the timestamp position in the load pattern coupling weight matrix is ​​extracted as the training weight within the fluctuation pattern. Simultaneously, the first and second causal correlation degrees corresponding to the cutoff time window at each fluctuation pattern's turning point are extracted as cross-pattern training weights, constructing a short-term training weight vector. It should be further noted that in this embodiment, the cross-pattern training uses the left and right boundaries of each cutoff time window as the input starting points for two different fluctuation patterns. Combining the load differential information corresponding to both sides of the fluctuation pattern's turning point and the cross-pattern training weights, synchronous sliding input training is performed from the left boundary to the fluctuation pattern's turning point and from the right boundary to the fluctuation pattern's turning point, respectively, according to the input length of the first sliding time window for the corresponding fluctuation pattern interval. Intra-mode training is the process of inputting training from left to right according to the corresponding fluctuation mode interval. The cross-mode training method in this embodiment takes the turning point of the fluctuation mode as the core. By clarifying the different fluctuation modes of the left and right boundaries, extracting the corresponding load differential information and cross-mode training weights, dynamically determining the sliding window length and bidirectional synchronous sliding input training, it accurately captures the dynamic correlation between load and temperature and electricity price during mode switching. It effectively solves the problem of insufficient feature characterization at the mode switching point in traditional training, greatly improves the prediction accuracy of power load in mode switching scenarios, and provides reliable technical support for the accurate formulation of enterprise power production and operation plans and the effective reduction of energy costs.

[0103] The first sliding time window for the corresponding fluctuation pattern interval is used as the ratio of the length of each identified fluctuation pattern interval to the standard deviation of the corresponding interval. The training input sequence set is obtained based on the load differential information combined with the length of the first sliding time window.

[0104] Based on the training input sequence set, short-term training weight vector, and gradient boosting tree, training is performed within the fluctuation mode interval and across modes to obtain the trained short-term prediction sub-model, and the first load prediction result and the first prediction error corresponding to the demand prediction time length are output.

[0105] It should be further explained that the construction process of the long-term prediction sub-model in this embodiment includes:

[0106] Obtain long-term forecast demand, load differential information, temperature and electricity price information for at least one year, and extract the corresponding forecast timestamp information, long-term forecast duration, current load differential information, temperature fluctuation sequence and electricity price fluctuation sequence;

[0107] Based on the predicted timestamp information and the current temperature fluctuation sequence and electricity price fluctuation sequence, combined with the correlation coupling feature and vector error correction model, the system is trained in the manner of training within the fluctuation mode interval and cross-mode training, and outputs the second load prediction result and the second prediction error.

[0108] While making long-term forecasts, synchronous fluctuation patterns are identified based on the second load forecast results obtained during the forecasting process. The trained short-term forecast sub-model is then called to perform different mode load forecasts for each preset short-term time length under the long-term forecast time length, so as to obtain the synchronous first load forecast results under the second load forecast results.

[0109] Based on the synchronous first load forecast result and the second load forecast result under the corresponding preset short-term time length, the forecast coupling error corresponding to the long-term forecast is obtained.

[0110] When the prediction coupling error is 0 and the second prediction error meets the preset long-term error threshold, the trained long-term prediction sub-model is obtained. The long-term error threshold is determined by statistically analyzing the historical prediction errors of the short-term prediction sub-model at the corresponding preset short-term time length, such as the mean error, variance error, and confidence interval. This is combined with the total number of preset short-term time lengths included within the long-term prediction time length. The short-term error statistics are adjusted according to the cumulative pattern of the short-term error statistics, such as multiplying the mean short-term error by the square root of the number of short-term segments. Finally, the error value obtained by dividing the adjusted error statistics by the number of short-term segments is used as a reference value for the long-term error threshold.

[0111] This application constructs a comprehensive power load forecasting technology system encompassing multi-mode accurate identification, multi-factor coupled quantification, and long-term / short-term collaborative verification. This system systematically addresses the core pain points of traditional power load forecasting, such as insufficient capture of fluctuation mode type conversion features corresponding to load fluctuation sequences, coarse characterization of the correlation between multiple influencing factors like temperature and electricity prices, disconnect between long-term and short-term forecast accuracy, and the tendency for error accumulation in long-term forecasts as the time scale expands. The underlying principle is derived layer by layer from three dimensions: feature processing, model construction, and collaborative optimization. In the feature processing dimension, the standardized load difference sequence is first deeply analyzed using a dynamic time-warping clustering algorithm to accurately identify… This study distinguishes between stationary fluctuation patterns with a difference result of 0, periodic fluctuation patterns with regular repetition, and transient fluctuation patterns with sudden changes, and labels each pattern with a corresponding time interval to achieve a refined classification of load fluctuation patterns. Then, based on factor analysis algorithms, the core features of each pattern—including the stationarity score and interval length of stationary patterns, the main period length and period intensity of periodic patterns, and the abrupt change amplitude and duration of transient patterns—are integrated with the concurrent temperature series, temperature difference series, electricity price series, and electricity price fluctuation series to construct a variable matrix specific to each fluctuation pattern interval. This matrix is ​​then standardized to eliminate differences in dimensions and magnitudes. Then, by calculating the correlation coefficient matrix between variables, common factors with eigenvalues ​​greater than 1 and a cumulative contribution rate of not less than 85% are extracted to determine the loading values ​​of temperature-related variables and electricity price-related variables on the common factors. Finally, the contribution of temperature and electricity price to the fluctuations in different mode intervals is calculated. Then, the preliminary correlation between temperature fluctuations, electricity price fluctuations, and load fluctuations is calculated by mode using a causal association analysis algorithm (correlation coefficient is calculated for stationary modes, similarity of periodic features is calculated for periodic modes, and matching degree of transient features is calculated for transient modes). Combined with Granger causality tests, the strength of the causal relationship among the three is verified and determined, forming the first causal correlation degree and the second causal correlation degree. Then, using a convolutional attention network, the load pattern coupling weight matrix, temperature fluctuation sequence, electricity price fluctuation sequence, and the corresponding fluctuation pattern type encoding of the load fluctuation sequence (stationary mode is denoted as 1, periodic mode as 2, and transient mode as 3) are instantaneously aligned by timestamp and multi-dimensional feature fusion is performed to generate correlation coupling features that can accurately characterize the degree of influence of temperature and electricity price on the fluctuation pattern type corresponding to the load fluctuation sequence. This process is progressive from pattern division to factor quantification, causal verification, and feature fusion, providing the prediction model with high-quality input data that has both pattern recognition and factor correlation, avoiding the prediction bias caused by traditional single features.In terms of model construction, the short-term prediction sub-model uses gradient boosting trees as its core algorithm and innovatively introduces a dual-track training mechanism of intra-mode training and cross-mode training. Intra-mode training follows the order of the fluctuation mode type intervals corresponding to the identified load fluctuation sequences, inputting load difference information and the fluctuation contribution of the corresponding modes into the model segment by segment according to the time series, ensuring sufficient learning of the load fluctuation pattern under a single mode. Cross-mode training takes the intercept window corresponding to the turning point of the fluctuation mode as the core, using the left boundary of the window as the input starting point of the previous mode and the right boundary as the input starting point of the next mode. Combining the load difference information on both sides of the turning point with the first causal correlation and the second causal correlation, synchronous sliding input training is performed from the left and right boundaries towards the turning point according to the length of the first sliding time window of the corresponding mode interval. This specifically solves the problem that traditional single-direction training is insufficient in characterizing the features at the mode transition, ensuring that short-term prediction can accurately capture key fluctuation details such as transient changes and cycle switching. The long-term forecast sub-model is based on a vector error correction model. Its core innovation lies in the long-short-term collaborative verification mechanism: When making long-term forecasts, the long-term forecast time range is first divided into several continuous and non-overlapping short-term time periods according to a preset short-term time length, ensuring that each short-term time period can cover the complete load fluctuation sub-pattern. Then, the trained short-term forecast sub-model is synchronously invoked to predict the load fluctuations in each short-term time period, generating a synchronous first load forecast result that is strictly aligned with the long-term forecast timestamp. This result is then compared with the second load forecast result output by the long-term forecast sub-model on a time-segment basis, and the prediction coupling error between the two is calculated. When the prediction coupling error is 0, it indicates that the long-term forecast is completely consistent with the short-term high-precision forecast in the fine-grained time dimension, effectively avoiding the problems of distortion of the fluctuation pattern type characteristics and accumulation of local errors caused by the amplification of the time scale in traditional long-term forecasts. At the same time, a preset second prediction error threshold is used for dual constraints, ensuring that the long-term forecast can fit the overall load change trend and that it conforms to the fine-grained law of load fluctuation in each short-term time period, achieving deep synergy between the accuracy of long-term forecasts and the accuracy of short-term forecasts.Through precise coupling of multiple modes and factors in the feature processing dimension and long-term and short-term collaborative verification in the model construction dimension, the system proposed in this application can maintain stable high-precision prediction performance under different load fluctuation scenarios: during stable fluctuation periods, the stability score and intra-mode training ensure the stability of prediction results; during periodic fluctuation periods, features such as main cycle length and cycle intensity are used in conjunction with periodic mode branch training to fit the regular fluctuations of the load; during transient fluctuation periods, features such as mutation amplitude and mutation steepness are used in conjunction with cross-mode training to accurately capture sudden changes in the load, and the long-term prediction accuracy does not decrease with the expansion of the prediction time scale. This provides reliable technical support for enterprises to formulate refined power production and operation plans, optimize energy dispatch schemes, and reduce energy consumption costs, significantly improving the adaptability, stability, and accuracy of power load prediction in complex and variable scenarios, and has important practical significance for promoting intelligent dispatch and efficient energy utilization in power systems.

[0112] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A power load forecasting system, characterized in that, include: The acquisition module is used to acquire load parameter sequences according to a preset sampling period length and combine them with time series algorithms to obtain correlation and coupling characteristics; The correlation coupling feature is used to characterize the degree of influence of seasonal factors, meteorological factors and electricity price fluctuation factors on load changes; The first prediction module is used to obtain the first load prediction result and the first prediction error, which characterize the short-term prediction, based on the load parameter sequence, a preset first sliding time window, and related coupling characteristics, through a preset short-term prediction sub-model. The second prediction module is used to obtain a second load prediction result and a second prediction error that characterize long-term prediction by combining the load parameter sequence with a second preset time window and the associated coupling characteristics through a preset long-term prediction sub-model. The online coupling module performs online coupling training and prediction based on the short-term prediction sub-model, the long-term prediction sub-model, the integrated framework, and the long-term and short-term coupling loss functions constructed by combining the first prediction error, the second prediction error, and the prediction coupling errors corresponding to the first load prediction result and the second load prediction result within the first sliding time window, to obtain the corresponding first load prediction result or second load prediction result.

2. The power load forecasting system as described in claim 1, characterized in that, The length of the preset sampling period is greater than the length of the second preset time window; The second preset time window length is greater than the first sliding time window length; The process of obtaining the associated coupling features includes: Collect temperature and electricity price data for the corresponding time periods of the load demand time series, and perform standardized preprocessing. Based on the processed load demand time series combined with the differential algorithm, a load differential series is obtained; Based on the load difference sequence combined with the dynamic time warping clustering algorithm and the preset fluctuation pattern types, fluctuation pattern similarity identification and clustering are performed to obtain fluctuation pattern clustering results; the fluctuation pattern types include stationary fluctuation patterns with a difference result of 0, periodic fluctuation patterns, and transient fluctuation patterns. Based on the load differential sequence under the steady fluctuation mode, the length of each steady fluctuation mode interval and the steadyness score are extracted; the steadyness score is equal to 1 minus the number of abnormal fluctuation points sampled within the steady fluctuation mode interval divided by the total number of sampling points.

3. The power load forecasting system as described in claim 2, characterized in that, The process of obtaining the associated coupling features also includes: Based on the load differential sequence under each cycle fluctuation mode and combined with Fourier transform, a frequency domain feature set under each cycle fluctuation mode is extracted; the frequency domain feature set includes the main cycle length and cycle intensity; the process of obtaining the cycle intensity includes: performing Fourier transform on the load differential sequence to obtain a frequency domain representation, identifying the frequency component with the largest amplitude as the main frequency and obtaining its amplitude, calculating the sum of squares of the amplitudes of all frequency components as the total signal energy, and taking the ratio of the amplitude of the main frequency to the total signal energy as the cycle intensity; Simultaneously, based on the load differential sequence under each cycle fluctuation mode and combined with statistical analysis algorithms, the time-domain feature set corresponding to each cycle fluctuation mode interval is obtained; the time-domain feature set includes the average rise or fall rate, peak or trough duration, fluctuation amplitude, and phase consistency characteristics. Based on the load differential sequence of each transient fluctuation mode and combined with statistical algorithms, a transient time-domain feature set is obtained; the transient time-domain feature set includes the mutation amplitude, mutation duration, mutation direction and mutation steepness.

4. The power load forecasting system as described in claim 3, characterized in that, The process of obtaining the associated coupling features also includes: Based on the stability score of each steady fluctuation mode in continuous time, the frequency domain feature set and time domain feature set of each periodic fluctuation mode, and the transient time domain feature set of each transient fluctuation mode, combined with the temperature sequence and corresponding temperature difference sequence, electricity price sequence and electricity price fluctuation sequence under the same continuous time, the contribution of temperature and electricity price variables to the fluctuation in each fluctuation mode interval is obtained by factor analysis algorithm. Based on the load differential sequence over continuous time, with the turning point of each fluctuation mode as the center of a preset intercept time window, and the standard deviation of the load differential sequences corresponding to both sides of the turning point of the fluctuation mode as the length of the intercept time window, the cross-modal load differential sequence of the preset intercept time window length is intercepted; the standard deviation length is specifically: ,in The length of the data truncation within the i-th fluctuation mode interval of the time window. To determine the data truncation length within the (i+1)th fluctuation pattern interval of the time window, and The standard deviations are, in order, the magnitudes of the standard deviations of the i-th fluctuation pattern interval and the (i+1)-th fluctuation pattern interval; and These are the interval lengths of the i-th fluctuation pattern interval and the (i+1)-th fluctuation pattern interval, respectively.

5. The power load forecasting system as described in claim 4, characterized in that, The process of obtaining the associated coupling features also includes: Based on the cross-modal load differential sequence combined with the temperature differential sequence and electricity price fluctuation sequence in the corresponding time period, the first causal correlation degree of temperature fluctuation on load fluctuation and the second causal correlation degree of electricity price fluctuation on load fluctuation in the corresponding time period are obtained through causal correlation analysis algorithm. The contribution of temperature and electricity price variables to the fluctuation within each fluctuation mode interval is standardized by combining the first causal correlation between temperature fluctuation and load fluctuation and the second causal correlation between electricity price fluctuation and load fluctuation in the corresponding time period, and then arranged according to the timestamp of the corresponding fluctuation mode interval to obtain the load mode coupling weight matrix. Based on the load pattern coupling weight matrix, combined with the temperature fluctuation sequence, the electricity price fluctuation sequence, and the fluctuation pattern type corresponding to the load fluctuation sequence, a convolutional attention network is used for convolutional coupling to obtain the correlation coupling features that characterize the degree of influence of temperature and electricity price on the fluctuation pattern type corresponding to the load fluctuation sequence.

6. The power load forecasting system as described in claim 5, characterized in that, The construction process of the short-term prediction sub-model includes: Obtain short-term load forecast demand and load differential information, temperature and electricity price information for a preset short-term time length, and extract the corresponding demand forecast time length, forecast timestamp information, and current temperature fluctuation sequence and electricity price fluctuation sequence. Based on the demand forecast time length, forecast timestamp information, load differential information, temperature fluctuation sequence and electricity price fluctuation sequence combined with the aforementioned correlation and coupling features, a pattern recognition algorithm is used to identify the fluctuation patterns contained within the demand forecast time length, the corresponding pattern interval length, the fluctuation pattern turning point, and the timestamp position of each current pattern interval length in the load pattern coupling weight matrix. The fluctuation contribution of each identified fluctuation pattern at the timestamp position in the load pattern coupling weight matrix is ​​extracted as the training weight within the fluctuation pattern. At the same time, the first causal correlation degree and the second causal correlation degree corresponding to the intercept time window at the turning point of each fluctuation pattern are extracted as cross-mode training weights to construct a short-term training weight vector.

7. The power load forecasting system as described in claim 6, characterized in that, The construction process of the short-term prediction sub-model also includes: The first sliding time window for the corresponding fluctuation pattern interval is used as the ratio of the length of each identified fluctuation pattern interval to the standard deviation of the corresponding interval. The training input sequence set is obtained based on the load differential information combined with the length of the first sliding time window. Based on the training input sequence set, short-term training weight vector, and gradient boosting tree, training is performed within the fluctuation mode interval and across modes to obtain the trained short-term prediction sub-model, and the first load prediction result and the first prediction error corresponding to the demand prediction time length are output.

8. The power load forecasting system as described in claim 7, characterized in that, The construction process of the long-term prediction sub-model includes: Obtain long-term forecast demand, load differential information, temperature and electricity price information for at least one year, and extract the corresponding forecast timestamp information, long-term forecast time length, current load differential information, temperature fluctuation sequence and electricity price fluctuation sequence; Based on the predicted timestamp information and the current temperature fluctuation sequence and electricity price fluctuation sequence, combined with the correlation coupling feature and vector error correction model, the system is trained in the manner of training within the fluctuation mode interval and cross-mode training, and outputs the second load prediction result and the second prediction error.

9. A power load forecasting system as described in claim 8, characterized in that, The construction process of the long-term prediction sub-model also includes: While making long-term forecasts, synchronous fluctuation patterns are identified based on the second load forecast results obtained during the forecasting process. The trained short-term forecast sub-model is then called to perform different mode load forecasts for each preset short-term time length under the long-term forecast time length, so as to obtain the synchronous first load forecast results under the second load forecast results. Based on the synchronous first load forecast result and the second load forecast result under the corresponding preset short-term time length, the forecast coupling error corresponding to the long-term forecast is obtained. When the prediction coupling error is 0 and the second prediction error meets the preset long-term error threshold, the trained long-term prediction sub-model is obtained.

10. The power load forecasting system as described in claim 9, characterized in that, The cross-mode training takes the left and right boundaries of each intercepted time window as the input starting points of two different fluctuation modes. Combining the load differential information and cross-mode training weights corresponding to the two sides of the fluctuation mode inflection point, synchronous sliding input training is performed from the left boundary to the fluctuation mode inflection point and from the right boundary to the fluctuation mode inflection point, respectively, according to the input length of the first sliding time window of the corresponding fluctuation mode interval.

Images