A multi-element load short-term prediction system for power systems

Abstract

The application relates to the technical field of power management, and particularly discloses a multi-element load short-term prediction system for a power system. The application obtains power grid actual operation data sources and meteorological data sources through a multi-source data acquisition platform, pre-processes the data sources, extracts local fluctuation characteristics according to the power grid actual operation data sources and the meteorological data sources, and fuses to generate a composite data sequence. A load trend perception model is designed based on the operation data sequence and the composite data sequence to analyze the power load recovery trend. The power grid dispatching is responded according to the prediction result, and a feedback closed loop is formed to continuously correct and optimize the prediction and dispatching strategy. The system structure fuses multi-dimensional internal and external information, takes into account short-term fluctuations and long-term trends, improves the prediction accuracy, and significantly enhances the robustness and self-adaptability of the system.

Description

A multi-element load short-term forecasting system for power systems Technical Field

[0001] This invention relates to the field of power management technology, and more specifically, to a multi-variable load short-term forecasting system for power systems. Background Technology

[0002] During blizzard events, power system load changes exhibit a multi-stage and multi-characteristic dynamic process, posing a severe challenge to traditional forecasting methods. Existing literature (Yang Kai. Research on High Resilience of Microgrid Systems in Highway Service Areas [D]. Chang'an University, 2023. DOI:10.26976 / d.cnki.gchau.2023.000568.) applies the concept of resilient grids to microgrids, conducting relevant research on the assessment of the resilience performance of microgrid systems and how to improve their resilience performance, and providing a schematic diagram of the response of the resilient microgrid before and after the disturbance, as shown in Figure 2. Blizzards not only change meteorological conditions but also affect user behavior and industrial and commercial activities, thereby causing drastic fluctuations in the power load curve.

[0003] In the initial stages of a blizzard, the power system typically receives advance notice from meteorological departments and early warning systems. These warnings prompt users and industrial enterprises in some areas to take preventative measures, such as increasing backup power or adjusting production plans, leading to rapid increases or decreases in load. As the blizzard progresses, meteorological conditions deteriorate dramatically, with wind, snow, and low temperatures becoming the dominant variables. Simultaneously, user behavior is influenced by the environment, exhibiting complex and non-linear characteristics. The sustained low temperatures and heavy snowfall during a blizzard lead to concentrated electricity consumption by residents to maintain essential living conditions. Meanwhile, travel disruptions and restrictions on commercial activities also cause irregular fluctuations in industrial load.

[0004] In the later stages of a blizzard, as weather conditions gradually improve, transportation, industrial, and commercial activities begin to recover, and user electricity demand shows signs of recovery. At this time, some areas may experience a gradual decline in load due to the release of additional energy reserves stored during the blizzard; while other areas may maintain a high load level due to special dispatching needs during the recovery period. Load data during the recovery phase typically exhibits a complex dynamic across multiple time scales, including short-term fluctuations caused by system adjustments before and after the blizzard, as well as the gradual increase in economic activity during the long-term recovery process. This complex and variable signal characteristic requires predictive models to flexibly adjust parameters and integrate more multi-source information in real time. However, traditional methods relying on a single data source and static model parameters struggle to achieve real-time dynamic updates, and their prediction results often suffer from delays or distortions, failing to accurately reflect the load recovery trend in the later stages of a blizzard, thus posing certain risks to power grid operation and dispatching. To address these issues, a technical solution is proposed. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides a multi-dimensional load short-term forecasting system for power systems. By integrating internal and external multi-dimensional information and considering both short-term fluctuations and long-term trends in the power load architecture, it addresses the problem that traditional methods relying on a single data source and static model parameters are difficult to achieve real-time dynamic updates and cannot accurately reflect the load rebound trend after a blizzard. This not only improves forecast accuracy but also significantly enhances the system's robustness and adaptability, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A multi-source load short-term forecasting system for power systems includes a data acquisition and processing module, a feature extraction and fusion module, a multi-model fusion forecasting module, and a dispatch response and closed-loop control module. The data acquisition and processing module acquires actual power grid operating data sources and meteorological data sources through a multi-source data acquisition platform and preprocesses the data sources. The feature extraction and fusion module extracts local fluctuation features from the actual power grid operating data sources and meteorological data sources and fuses them to generate a composite data sequence. The multi-model fusion forecasting module designs a load trend perception model based on the operating data sequence and the composite data sequence to analyze the power load recovery trend. A first load trend perception unit extracts first local fluctuation features from the operating data sequence and second local fluctuation features from the meteorological data sequence to form a fusion vector. Based on the first local fluctuation features, the second local fluctuation features, and the fusion vector, a first load trend perception model is constructed to predict the first load trend perception index. The construction steps of the first load trend perception model include:

[0008] A fusion vector is constructed by concatenating the first and second data distribution features. , The first data distribution feature vector, This is the feature vector of the second data distribution;

[0009] Based on the first data change rate Second data change rate The first load trend perception model is constructed by constructing adjacent fusion vectors to predict the first load trend perception index. The average value of the absolute difference between the rate of change of the power grid operation data and various meteorological data is calculated in each time window. The average value is multiplied by the fusion vector in the window to obtain the first sensitive vector. The first sensitive vector between adjacent windows is compared to obtain the first load trend perception index.

[0010] As a further embodiment of the present invention, the data acquisition and processing module includes a multi-source data acquisition platform and a data preprocessing unit;

[0011] The multi-source data acquisition platform is used to acquire data from actual power grid operation data sources and meteorological data sources; the actual power grid operation data sources include historical load data and real-time load data; the meteorological data sources include temperature, humidity, snowfall, and wind speed.

[0012] The data preprocessing unit is used to normalize, correct, detect outliers, process missing data, and perform time alignment on the actual power grid operation data source and meteorological data source.

[0013] As a further embodiment of the present invention, the feature extraction and fusion module includes a feature extraction unit and a feature fusion unit; the feature extraction unit is used to perform time-domain analysis on the actual power grid operation data source and the meteorological data source respectively, and to extract local fluctuation features by obtaining continuous data sequences through sliding window technology; the local fluctuation features include a first local fluctuation feature and a second local fluctuation feature; the feature fusion unit is used to splice the operation data sequence and the meteorological data sequence to construct a composite data sequence.

[0014] As a further aspect of the present invention, the first local fluctuation characteristic is obtained by calculating the first data change rate and the first data distribution characteristic based on the operating data sequence; the first data distribution characteristic includes the mean, standard deviation, skewness, and peak value of the operating data sequence within the window; the first data change rate is obtained by calculating the change in the actual operating data of the power grid between two adjacent sampling points in the operating data sequence within the same window, and calculating the first data change rate based on the mean of the change; the formula for calculating the first data change rate is:

[0015] ;

[0016] In the formula: The rate of change of the first data within the k-th window. This represents the actual operating data of the (i+1)th power grid. For the actual operating data of the i-th power grid, The amount of data in the k-th window;

[0017] The formula for calculating the mean of the data sequence running within the window is:

[0018] ;

[0019] In the formula: The mean of the data sequence for the k-th window;

[0020] The formula for calculating the standard deviation is:

[0021] ;

[0022] In the formula: The standard deviation of the data sequence for the k-th window;

[0023] The formula for calculating the bias is:

[0024] ;

[0025] In the formula: The bias value of the data sequence for the k-th window;

[0026] The formula for calculating the peak value is:

[0027] ;

[0028] In the formula: Peak value of the data sequence for the k-th window

[0029] As a further aspect of the present invention, the second local fluctuation feature is obtained by calculating the second data change rate and the second data distribution feature based on the meteorological data sequence; the second data distribution feature includes the mean, standard deviation, skewness, and peak value of the meteorological data sequence within the window; the second data change rate is obtained by calculating the change in meteorological data between two adjacent sampling points in the running data sequence within the same window, and calculating the second data change rate based on the mean of the change. The formula for calculating the second data change rate is:

[0030] ;

[0031] In the formula: The rate of change of the second data within the k-th window. For the (i+1)th meteorological data, For the i-th meteorological data, Let be the amount of data in the k-th window.

[0032] As a further embodiment of the present invention, the multi-model fusion prediction module includes a first load trend sensing unit, a second load trend sensing unit, and a trend-sensitive feature analysis unit; the second load trend sensing unit is used to design a second load trend sensing model based on the fused composite data sequence to predict the second load trend sensing index; the trend-sensitive feature analysis unit is used to analyze the power load recovery trend based on the first load trend sensing index and the second load trend sensing index.

[0033] As a further aspect of the present invention, the step of the second load trend sensing unit in designing a second load trend sensing model and predicting a second load trend sensing index based on the fused composite data sequence includes:

[0034] By obtaining composite data sequences from adjacent time windows, the difference sequence is obtained by calculating the difference vector at each time point.

[0035] The second load trend perception model is constructed based on the difference sequence to predict the second load trend perception index. For each pair of adjacent time windows, the absolute difference between the i-th difference vector of the (k+1)-th window and the corresponding difference vector in the previous window is calculated. All these absolute differences are summed and then divided by the total number of time windows Q−1 to obtain the average difference between all adjacent windows.

[0036] As a further aspect of the present invention, the trend-sensitive feature analysis unit is used to analyze the power load recovery trend based on a first load trend sensing index and a second load trend sensing index. If the difference between the first load trend sensing index and the second load trend sensing index is... If the value is greater than or equal to a preset threshold, the power load recovery trend is stable; if the difference between the first load trend sensing index and the second load trend sensing index is greater than or equal to a preset threshold, the power load recovery trend is stable. If the load is below the preset threshold, the power load recovery trend will be unstable.

[0037] The technical effects and advantages of this invention, a multi-source load short-term forecasting system for power systems, are as follows: This invention acquires actual power grid operation data sources and meteorological data sources through a multi-source data acquisition platform, preprocesses the data sources, extracts local fluctuation characteristics based on the actual power grid operation data sources and meteorological data sources, and fuses them to generate composite data sequences. Based on the operation data sequences and composite data sequences, a load trend perception model is designed to analyze the power load recovery trend. The power grid dispatch is responded to according to the forecast results, and a feedback closed loop is formed to continuously correct and optimize the forecasting and dispatching strategies. The system architecture, which integrates internal and external multi-dimensional information and takes into account both short-term fluctuations and long-term trends, not only improves the forecasting accuracy but also significantly enhances the robustness and adaptability of the system. Attached Figure Description

[0038] Figure 1 is a comparison chart of actual load and predicted load provided by the present invention;

[0039] Figure 2 is a schematic diagram of the response of the elastic microgrid before and after a disturbance provided by the present invention;

[0040] Figure 3 is a heatmap of multi-source data correlation provided by the present invention;

[0041] Figure 4 is a schematic diagram of the structure of a multi-element load short-term forecasting system for power systems provided by the present invention. Detailed Implementation

[0042] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described technical solutions are only a part of this invention, not all of it. All other technical solutions obtained by those skilled in the art based on the technical solutions of this invention without inventive effort are within the scope of protection of this invention. Examples

[0043] Figure 4 is a schematic diagram of the structure of a multi-load short-term forecasting system for a power system provided by the present invention. As shown in the figure, the multi-load short-term forecasting system for a power system includes a data acquisition and processing module, a feature extraction and fusion module, a multi-model fusion forecasting module, and a dispatch response and closed-loop control module. The data acquisition and processing module is connected to the feature extraction and fusion module, the feature extraction and fusion module is connected to the multi-model fusion forecasting module, and the multi-model fusion forecasting module is connected to the dispatch response and closed-loop control module.

[0044] The data acquisition and processing module is used to acquire actual power grid operation data sources and meteorological data sources through a multi-source data acquisition platform, and to preprocess the data sources.

[0045] The feature extraction and fusion module is used to extract local fluctuation features based on the actual power grid operation data source and meteorological data source, and fuse them to generate a composite data sequence;

[0046] The multi-model fusion prediction module designs a load trend perception model based on operational data sequences and composite data sequences to analyze the power load recovery trend;

[0047] The dispatch response and closed-loop control module is used to respond to power grid dispatch based on the forecast results and form a feedback closed loop to continuously correct and optimize the forecast and dispatch strategies.

[0048] Specifically, the data acquisition and processing module includes a multi-source data acquisition platform and a data preprocessing unit;

[0049] The multi-source data acquisition platform is used to acquire data from actual power grid operation data sources and meteorological data sources; the actual power grid operation data sources include historical load data and real-time load data; the meteorological data sources include temperature, humidity, snowfall, and wind speed.

[0050] The data preprocessing unit is used to normalize, correct, detect outliers, process missing data, and perform time alignment on the actual power grid operation data source and meteorological data source.

[0051] Specifically, the feature extraction and fusion module includes a feature extraction unit and a feature fusion unit; the feature extraction unit and the feature fusion unit are connected.

[0052] The feature extraction unit is used to perform time-domain analysis on the actual power grid operation data source and the meteorological data source respectively, and to extract local fluctuation features by obtaining continuous data sequences through the sliding window technique;

[0053] Time-domain analysis is performed on the actual data sources of the power grid operation, and continuous operation data sequences are obtained through the sliding window technique. , For the actual operating data of the i-th power grid, For the nth actual power grid operation data within the kth window, the first local fluctuation characteristics are captured based on the operation data sequence; time-domain analysis is performed on the meteorological data source, and continuous meteorological data sequences are obtained through sliding window technology. , For the i-th meteorological data, For the nth meteorological data within the kth window, the second local fluctuation characteristics are captured based on the meteorological data sequence.

[0054] The feature fusion unit is used to splice together the operational data sequence and the meteorological data sequence to construct a composite data sequence.

[0055] By simultaneously collecting actual power grid operation data (including historical and real-time load) and meteorological data (temperature, humidity, snowfall, wind speed, etc.), a comprehensive understanding of the operation can be obtained, providing a more complete reflection of the power grid's operating status and significantly improving the accuracy of the prediction model. Employing a sliding window technique for time-domain analysis of different data, local fluctuation characteristics in both power grid operation and meteorological data are extracted. Feature fusion is then used to construct a composite data sequence, which helps capture subtle changes and trends in the load, thereby improving the sensitivity and accuracy of the prediction. By integrating the advantages of multiple prediction models, the limitations and errors that may exist in a single model can be reduced, enhancing the robustness of the overall prediction results. When facing complex and variable power grid operating environments and meteorological conditions, multi-model fusion can more stably reflect load trends. The dispatch response and closed-loop control module not only transforms the prediction results into dispatch commands but also feeds back actual operation data to the system for correcting and optimizing model parameters and dispatch strategies, achieving adaptive regulation. This closed-loop control mechanism improves the system's response capability to emergencies and abnormal data, further ensuring the safety and stability of power grid operation.

[0056] Specifically, the first local fluctuation characteristic is obtained by calculating the first data change rate and the first data distribution characteristic based on the operating data sequence; the first data distribution characteristic includes the mean, standard deviation, skewness, and peak value of the operating data sequence within the window; the first data change rate is obtained by calculating the change in the actual operating data of the power grid between two adjacent sampling points in the operating data sequence within the same window, and calculating the first data change rate based on the mean of the change; the formula for calculating the first data change rate is:

[0057] ;

[0058] In the formula: The rate of change of the first data within the k-th window. This represents the actual operating data of the (i+1)th power grid. For the actual operating data of the i-th power grid, The amount of data in the k-th window;

[0059] The formula for calculating the mean of the data sequence running within the window is:

[0060] ;

[0061] In the formula: The mean of the data sequence for the k-th window;

[0062] The formula for calculating the standard deviation is:

[0063] ;

[0064] In the formula: The standard deviation of the data sequence for the k-th window;

[0065] The formula for calculating the bias is:

[0066] ;

[0067] In the formula: The bias value of the data sequence for the k-th window;

[0068] The formula for calculating the peak value is:

[0069] ;

[0070] In the formula: The peak value of the data sequence for the k-th window.

[0071] Specifically, the second local fluctuation characteristic is obtained by calculating the second data change rate and the second data distribution characteristic based on the meteorological data sequence; the second data distribution characteristic includes the mean, standard deviation, skewness, and peak value of the meteorological data sequence within the window; the second data change rate is calculated by calculating the change in meteorological data between two adjacent sampling points in the running data sequence within the same window, and the second data change rate is calculated based on the mean of the change; the formula for calculating the second data change rate is:

[0072] ;

[0073] In the formula: The rate of change of the second data within the k-th window. For the (i+1)th meteorological data, For the i-th meteorological data, Let be the amount of data in the k-th window.

[0074] By calculating the rate of change of the first (or second) data using the difference between adjacent sampling points within a window, subtle changes in power grid operation data or meteorological data within a short period of time can be sensitively captured. This method can promptly reflect sudden trends in load or meteorological conditions, providing important information for subsequent load recovery trend analysis. By calculating the mean, standard deviation, skewness (reflecting data asymmetry), and peak value (reflecting the degree of peaks or flatness) of the data within the window, the distribution characteristics of the data are fully described. Such statistical information is of great significance for revealing the overall trend and local fluctuation characteristics of the data. Processing continuous data through a sliding window and calculating statistics such as the mean can, to some extent, smooth out noise introduced by measurement errors or random disturbances, making the subsequent model more robust in capturing trends. At the same time, statistical features such as standard deviation, skewness, and peak value can help identify anomalies in data distribution, providing a basis for detecting and correcting abnormal data and improving data quality. Using statistics and rates of change can compress the rich information in the original data sequence into several descriptive indicators, which not only reduces the complexity of data processing but also provides more refined and representative input features for subsequent multi-model fusion prediction. By calculating the corresponding local fluctuation characteristics of power grid operation data and meteorological data separately and then fusing them, the key dynamic characteristics of the two data sources can be organically combined, thereby improving the predictive model's ability to grasp comprehensive influencing factors.

[0075] Specifically, the multi-model fusion prediction module includes a first load trend sensing unit, a second load trend sensing unit, and a trend-sensitive feature analysis unit; the first load trend sensing unit and the second load trend sensing unit are respectively connected to the trend-sensitive feature analysis unit.

[0076] The first load trend sensing unit is used to extract the first local fluctuation features of the operational data sequence and the second local fluctuation features of the meteorological data sequence. Based on the first data distribution features and the second data distribution features, a fusion vector is formed by splicing them together. Based on the first local fluctuation features, the second local fluctuation features, and the fusion vector, a first load trend sensing model is constructed to predict the first load trend sensing index.

[0077] The second load trend sensing unit is used to design a second load trend sensing model based on the fused composite data sequence to predict the second load trend sensing index.

[0078] The trend-sensitive feature analysis unit is used to analyze the power load recovery trend based on the first load trend perception index and the second load trend perception index.

[0079] Specifically, a first load trend perception model is constructed based on the first data change rate, the second data change rate, and the fusion vector to predict the first load trend perception index. The construction steps of the first load trend perception model are as follows:

[0080] A fusion vector is constructed by concatenating the first and second data distribution features. , The first data distribution feature vector, This is the second data distribution feature vector. , ,in, The mean of the data sequence run in the k-th window. The standard deviation of the data sequence for the k-th window. The bias of the data sequence for running in the k-th window. The peak value of the data sequence for the k-th window. Let be the mean of the meteorological data sequence for the k-th window. Let be the standard deviation of the meteorological data series for the k-th window. The bias value of the meteorological data sequence for the k-th window. The peak value of the meteorological data sequence in the k-th window;

[0081] Based on the first data change rate Second data change rate The first load trend sensing model is constructed using adjacent fusion vectors to predict the first load trend sensing index. This is achieved by first calculating the average absolute difference between the rate of change of power grid operation data and various meteorological data within each time window, then multiplying this average by the fusion vector within the window to obtain the first sensitivity vector. The first sensitivity vectors between adjacent windows are then compared to obtain the first load trend sensing index. The formula for the first load trend sensing model is:

[0082] ;

[0083] In the formula: As the primary indicator for sensing load trends, Number of time windows This represents the number of data items in the meteorological data source. The rate of change of the first data in the (k+1)th time window. The rate of change of the second data item a in the (k+1)th time window. Let be the fusion vector for the (k+1)th time window. The rate of change of the first data in the k-th time window. The rate of change of the second data item a in the k-th time window. This is the fusion vector for the k-th time window;

[0084] It should be noted that the number of data items in the meteorological data source... The value is 4, representing temperature, humidity, snowfall, and wind speed, respectively.

[0085] The distribution characteristics of data within each time window are comprehensively described using mean, standard deviation, skewness, and peak value, allowing for a detailed portrayal of the local dynamics of power grid load and meteorological data. By calculating the changes between adjacent sampling points, subtle trends in data changes can be quickly captured, providing sensitive dynamic indicators for predicting power load recovery. A fusion vector is constructed by concatenating the statistical characteristics of power grid operation data and meteorological data, integrating information from both data sources while highlighting their respective characteristics, resulting in a more comprehensive overall information representation. For meteorological data, four dimensions—temperature, humidity, snowfall, and wind speed—are considered, enabling the model to better reflect the complex impact of meteorological factors on load trends and reducing the risk of misjudgment based on a single factor. Sensitive indicators are constructed based on the rate of change of the first and second data sets and the fusion vector. By comparing the changes in sensitive vectors between consecutive windows, shifts in load trends are effectively captured. Modeling the fused composite data sequence supplements the prediction dimensions, making the overall prediction more comprehensive and accurate in responding to data changes. Combining the two sets of indicators, the power load recovery trend is further analyzed, ensuring that the prediction results are more credible and actionable. By constructing a first load trend perception index using the difference between adjacent fusion vectors and the rate of change, this method effectively suppresses noise interference on individual data fluctuations, making the prediction model more robust in the face of complex and ever-changing environments. By comparing the index changes across different time windows, the model can update its perception of load trends in real time and dynamically correct for future load recovery trends, thereby achieving long-term stable operation and continuous adaptive optimization.

[0086] Specifically, the second load trend sensing unit is used to design a second load trend sensing model based on the fused composite data sequence to predict the second load trend sensing index:

[0087] By obtaining composite data sequences from adjacent time windows, the difference sequence is obtained by calculating the difference vector at each time point. , , Let i be the difference vector of the k-th time window. For the i-th running data in the (k+1)-th time window, For the i-th running data in the k-th time window, For the i-th meteorological data in the (k+1)-th time window, This refers to the i-th meteorological data point within the k-th time window.

[0088] The second load trend perception model is constructed based on the difference sequence to predict the second load trend perception index. For each pair of adjacent time windows (from the kth to the (k+1th), there are a total of Q−1 pairs), the absolute difference between the i-th difference vector of the (k+1th)th window and the corresponding difference vector in the previous window is calculated. All these absolute differences are summed and then divided by the number of time window pairs Q−1 to obtain the average difference between all adjacent windows.

[0089] The formula for the second load trend sensing model is:

[0090] ;

[0091] In the formula: This is the second indicator for sensing load trends. This represents the i-th difference vector within the (k+1)-th time window. Let i be the difference vector of the k-th time window. This represents the number of time windows.

[0092] Specifically, the trend-sensitive feature analysis unit is used to analyze the power load recovery trend based on the first load trend sensing index and the second load trend sensing index. If the difference between the first load trend sensing index and the second load trend sensing index is... If the value is greater than or equal to a preset threshold, the power load recovery trend is stable; if the difference between the first load trend sensing index and the second load trend sensing index is greater than or equal to a preset threshold, the power load recovery trend is stable. If the load is below the preset threshold, the power load recovery trend will be unstable.

[0093] Figure 1 is a comparison chart of actual load and predicted load provided by the present invention. The heat map involves multiple key variables such as power load, temperature, humidity and wind speed. The cells with intersecting rows and columns show the correlation coefficient between the corresponding two variables. The color from light to dark usually indicates from low correlation to high correlation. The closer the value is to 1, the stronger the positive correlation. The closer it is to -1, the stronger the negative correlation. The correlation coefficient at the diagonal position (variable and itself) is 1.

[0094] Figure 3 is a heatmap of multi-source data correlation provided by this invention; the horizontal axis (time) shows the time series from early morning to dawn, morning, and even the whole day; the vertical axis (value) is usually the load power (e.g., kilowatts or megawatts), used to represent the size of the power load within a unit time period. The blue curve (or actual curve) represents the actual change of the measured power load over time; the green curve (or predicted curve) represents the load prediction result obtained based on the multivariate prediction model (combining meteorological and other multi-source data). By comparing the two curves, the accuracy of the model prediction can be intuitively evaluated: when the two curves are close together, it indicates that the short-term load prediction effect is good; if the deviation is significant, the reasons in the model or data need to be analyzed. In the figure, at 08:00, the actual load is about 901, while the predicted load is about 934, so the error is about tens of units, which may indicate that there is a certain deviation in the model, but the overall trend is still consistent.

[0095] By constructing a difference vector using composite data sequences from adjacent windows, the model accurately quantifies the changes in power grid operation data and meteorological data over continuous time periods, enabling it to keenly capture subtle dynamics during load recovery. By calculating and averaging the absolute differences between adjacent windows, a specific trend-sensing index R2 is obtained, providing an intuitive numerical basis for subsequent judgment of load recovery trends. This approach not only considers individual changes in operation and meteorological data but also reflects the dynamic relationship between them through composite data sequences, allowing the model to fully utilize multi-source information and comprehensively consider the impact of various factors on load fluctuations. The difference calculation method helps detect the consistency of dynamic changes between adjacent time windows, thus providing a supplementary perspective for load recovery trend analysis and compensating for local abrupt changes or fluctuations that might be overlooked by a single index. The trend-sensitive feature analysis unit compares the second load trend-sensing index R2 with the first load trend-sensing index R1, using a preset threshold to determine the magnitude of their difference, objectively distinguishing between stable and unstable load recovery trends. Example

[0096] The continuous data was divided into 5 time windows (Q=5), and the number of data items in the meteorological data source was fixed at 4 (A=4, representing temperature, humidity, snowfall, and wind speed, respectively). Within each time window, the first data change rate of the power grid operation data and the second data change rate of each meteorological data item were calculated. Simultaneously, the distribution characteristics (mean, standard deviation, skewness, and peak value) of the power grid and meteorological data within the window were obtained using sliding window statistics. The two feature sets were then concatenated to form a fusion vector, and the L2 norm of the fusion vector was used as the scalar factor F. kFinally, the average absolute difference calculated within each window is multiplied by the fusion factor of that window to obtain the first sensitivity vector. The first sensitivity vectors between adjacent windows are compared to obtain the first load trend perception index. For the second load trend perception model, the second load trend perception index is calculated by obtaining the difference vector between consecutive windows of the composite data sequence for each time window. Table 1 below shows the key parameters for each time window and the calculation of the fusion sensitivity vector.

[0097] Table 1. Calculation of key parameters and fusion sensitivity vector for each time window.

[0098] Based on the first load trend perception index in the table above, the difference in sensitivity vector values ​​between adjacent windows is calculated as follows:

[0099] Window 1 and 2: |16.97−12.22|=4.75;

[0100] Window 2 and 3: |5.50−16.97|=11.47;

[0101] Window 3 and 4: |5.70−5.50|=0.20;

[0102] Window 4 and 5: |5.60−5.70|=0.10;

[0103] Then there is

[0104] ;

[0105] By calculating the difference vector at each time point in the composite data sequence of each window, the average absolute differences between adjacent windows are obtained as 0.8, 1.2, 0.6, and 0.5, respectively. Therefore:

[0106] ;

[0107] According to the judgment rules of the trend-sensitive feature analysis unit: if |R1−R2|≥ preset threshold (e.g., 10), the power load recovery trend is considered to be stable; if |R1−R2|< preset threshold, the trend is considered to be unstable.

[0108] Calculated

[0109] ;

[0110] Since 15.74 is greater than the preset threshold of 10, it can be determined that the power load recovery trend is stable.

[0111] This invention utilizes a multi-source data acquisition platform to simultaneously acquire actual power grid operation data and meteorological data (such as temperature, humidity, snowfall, and wind speed). This allows the model to go beyond relying on a single data source and comprehensively consider the impact of internal and external factors on load. Combining historical load data with real-time data enables the capture of long-term trends and short-term fluctuations in power grid operation, further improving data accuracy and dynamic response capabilities. A sliding window technique is employed to perform time-domain analysis on both power grid and meteorological data, extracting first and second local fluctuation features to better capture dynamic changes within the data. The statistical distribution characteristics (mean, standard deviation, skewness, and peak value) of power grid and meteorological data are concatenated to form a fusion vector, allowing the model to understand the system state from multiple dimensions and leverage the synergistic effect of internal and external data in subsequent predictions. The constructed first and second load trend perception models focus on changes in power grid data and composite data sequences, respectively. Each model has its strengths; through the fusion of multiple models, they can complement each other, reducing the bias that a single model might introduce. By comparing and statistically analyzing the sensitivity vector and composite data difference vector within adjacent time windows, indicators are constructed respectively, and the difference between the two is used as a trend-sensitive feature to help accurately determine whether the load recovery trend is stable.

[0112] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0113] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-variable load short-term forecasting system for power systems, comprising a data acquisition and processing module, a feature extraction and fusion module, a multi-model fusion forecasting module, and a dispatch response and closed-loop control module; characterized in that, The data acquisition and processing module is used to acquire actual power grid operation data sources and meteorological data sources through a multi-source data acquisition platform, and to preprocess the data sources; the feature extraction and fusion module is used to extract local fluctuation features based on the actual power grid operation data sources and meteorological data sources, and to fuse them to generate composite data sequences. The multi-model fusion prediction module designs a load trend perception model based on operational data sequences and composite data sequences to analyze the power load recovery trend; the first load trend perception unit is used to extract the first local fluctuation features of the operational data sequence and the second local fluctuation features of the meteorological data sequence to form a fusion vector. The first local fluctuation feature is obtained by calculating the first data change rate and the first data distribution feature based on the operational data sequence; the second local fluctuation feature is obtained by calculating the second data change rate and the second data distribution feature based on the meteorological data sequence. A first load trend perception model is constructed based on the first data change rate, the second data change rate, and the fusion vector to predict the first load trend perception index. The construction steps of the first load trend perception model include: constructing a fusion vector by concatenating the first data distribution characteristics and the second data distribution characteristics. ， The first data distribution feature vector, The second data distribution feature vector; based on the first data change rate Second data change rate The first load trend perception model is constructed using adjacent fusion vectors to predict the first load trend perception index. This is achieved by first calculating the average absolute difference between the rate of change of the power grid operation data and various meteorological data within each time window, then multiplying the average value by the fusion vector within the window to obtain the first sensitivity vector. The first sensitivity vectors between adjacent windows are compared to obtain the first load trend perception index. The feature extraction and fusion module includes a feature extraction unit and a feature fusion unit. The feature extraction unit performs time-domain analysis on the actual power grid operation data source and the meteorological data source, respectively, and extracts local fluctuation features from continuous data sequences using sliding window technology. These local fluctuation features include first local fluctuation features and second local fluctuation features. The feature fusion unit splices the operation data sequence and the meteorological data sequence to construct a composite data sequence. (First Bureau) The first data fluctuation characteristic is obtained by calculating the first data change rate and the first data distribution characteristic based on the operating data sequence. The first data distribution characteristic includes the mean, standard deviation, skewness, and peak value of the operating data sequence within the window. The first data change rate is obtained by calculating the change in the actual operating data of the power grid between two adjacent sampling points in the operating data sequence within the same window, and calculating the first data change rate based on the mean of the change. The second local fluctuation characteristic is obtained by calculating the second data change rate and the second data distribution characteristic based on the meteorological data sequence. The second data distribution characteristic includes the mean, standard deviation, skewness, and peak value of the meteorological data sequence within the window. The second data change rate is obtained by calculating the change in the meteorological data between two adjacent sampling points in the operating data sequence within the same window, and calculating the second data change rate based on the mean of the change.

2. The multi-source load short-term forecasting system for power systems according to claim 1, characterized in that, The data acquisition and processing module includes a multi-source data acquisition platform and a data preprocessing unit; the multi-source data acquisition platform is used to acquire data from actual power grid operation data sources and meteorological data sources; the actual power grid operation data sources include historical load data and real-time load data; Meteorological data sources include temperature, humidity, snowfall, and wind speed; The data preprocessing unit is used to normalize, correct, detect outliers, process missing data, and perform time alignment on the actual power grid operation data source and meteorological data source.

3. The multi-source load short-term forecasting system for power systems according to claim 1, characterized in that, The multi-model fusion prediction module includes a first load trend sensing unit, a second load trend sensing unit, and a trend-sensitive feature analysis unit; The second load trend sensing unit is used to design a second load trend sensing model based on the fused composite data sequence to predict the second load trend sensing index. The trend-sensitive feature analysis unit is used to analyze the power load recovery trend based on the first load trend perception index and the second load trend perception index.

4. A multi-element load short-term forecasting system for power systems according to claim 3, characterized in that, The steps of the second load trend sensing unit for designing a second load trend sensing model and predicting the second load trend sensing index based on the fused composite data sequence include: obtaining composite data sequences of adjacent time windows, calculating the difference vector at each time point to obtain the difference sequence; constructing a second load trend sensing model to predict the second load trend sensing index based on the difference sequence; for each pair of adjacent time windows, calculating the absolute difference between the i-th difference vector of the (k+1)-th window and the corresponding difference vector in the previous window; summing all these absolute differences and then dividing by the total number of time windows Q−1 to obtain the average difference between all adjacent windows.

5. A multi-element load short-term forecasting system for power systems according to claim 3, characterized in that, The trend-sensitive feature analysis unit is used to analyze the power load recovery trend based on the first load trend sensing index and the second load trend sensing index. If the difference between the first load trend sensing index and the second load trend sensing index... If the load is greater than or equal to the preset threshold, the power load recovery trend is stable. If the difference between the first load trend perception index and the second load trend perception index If the load is below the preset threshold, the power load recovery trend will be unstable.

Images