A power load prediction method based on fractional calculus control and graph neural network fusion

Abstract

The application discloses a power load prediction method based on fractional calculus control and graph neural network fusion, comprising the following steps: based on historical load data and multi-source factor data, a long short-term memory network is used to extract a time series preliminary trend feature vector; if the regional influence intensity exceeds a preset threshold, the sudden demand information is fused to determine an enhanced trend vector; an optimized trend representation data is obtained by combining a convolutional neural network, a weighted trend model is constructed by fusing the weight distribution of multi-source factors through an attention mechanism; the backtracking adjustment parameter and data quality filtering are fused, and according to the stable load prediction result output by the final prediction model, a coordinated mapping between multiple regions and a comprehensive power load optimization scheme are generated. The application fuses multi-dimensional features and a dynamic iterative updating mechanism, forms a dynamic adaptive system, significantly improves the accuracy of load prediction and power supply stability under the influence of multiple complex factors, and realizes the optimal allocation of power resources.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a power load forecasting method based on the fusion of fractional-order calculus control and graph neural networks. Background Technology

[0002] Electricity load forecasting is a core component of power system operation and management, directly impacting socio-economic stability, energy efficiency, and the achievement of environmental protection goals. Accurate load forecasting can prevent power outages caused by insufficient supply or resource idleness due to oversupply, thereby reducing economic losses and optimizing generation dispatch.

[0003] However, existing technologies struggle to cope with volatile environments and demand fluctuations, leading to frequent forecasting errors and hindering the reliability and sustainability of power systems. On one hand, existing methods often overlook the underlying spatiotemporal patterns behind load changes when dealing with diverse demands in electricity load forecasting. For example, models struggle to capture nonlinear patterns during peak-hour electricity demand surges in cities or seasonal agricultural load fluctuations in rural areas, especially when historical data is scarce or contaminated by noise, causing a sharp decline in forecast reliability. This deficiency manifests as amplified hourly deviations in short-term forecasts such as intraday dispatching, resource misallocation in medium-term forecasts such as weekly planning, and amplified uncertainty in long-term forecasts such as monthly planning. This results in inaccurate dispatching decisions when the system faces unforeseen events such as extreme weather or peak demand during holidays, triggering a chain reaction.

[0004] On the other hand, existing methods face the challenge of accurately identifying load trend characteristics and integrating the coupling of multiple factors. Load trends often manifest as implicit patterns in historical data, such as periodic fluctuations between weekdays and weekends, or exponential increases in air conditioning load during hot weather. These characteristics are driven by multiple factors, including time series data, meteorological variables, and economic activities. If these trends cannot be extracted from noisy data, it is difficult to predict future paths. For example, if the load in an industrial park suddenly drops due to factory maintenance, and the model does not correlate with the compensatory increase in surrounding residential areas, it may underestimate overall demand, leading to delays in the activation of backup power.

[0005] Furthermore, the interactions between different power grid regions further complicate the problem. Loads in different regions are not isolated. For example, high temperatures in North China leading to a surge in air conditioning load may affect load distribution in East China through inter-regional power transmission. If forecasts ignore this spatial dependence, localized overloads or global inefficiencies will be inevitable. In scenarios with low data quality or data acquisition delays, this regional interconnectedness becomes even more difficult to track. For instance, power grids in remote mountainous areas are affected by fluctuations in upstream hydropower, leading to distorted load forecasts for downstream cities, resulting in uneven energy delivery and increased environmental pollution.

[0006] The aforementioned issues highlight the technical contradictions in business operations. While pursuing high-precision forecasting, it is essential to balance adaptability across multiple time scales, robustness against data scarcity, and the capture of regional spatiotemporal correlations. Short-term forecasts require responding to instantaneous fluctuations, medium-term forecasts need to integrate seasonal patterns, and long-term forecasts require predicting macroeconomic trends. However, existing technologies often sacrifice other aspects for optimization in one dimension, resulting in overall low efficiency. If these technical bottlenecks cannot be overcome, the power system will continue to face dispatch risks, impacting grid resilience and the progress of green transformation. Summary of the Invention

[0007] This invention aims to provide a power load forecasting method based on the fusion of fractional-order calculus control and graph neural networks. It can propose a comprehensive optimization scheme to address complex business scenarios in power load forecasting, such as the dynamic correlation of multiple factors, the intensity of regional effects, sudden changes in demand, and fluctuations in environmental impacts, thereby improving the accuracy of load forecasting and the stability of power supply under the influence of multiple complex factors.

[0008] This invention provides a power load forecasting method based on the fusion of fractional-order calculus control and graph neural networks, the method comprising: Based on historical load data and multi-source factor data, a preliminary trend feature vector of the time series is extracted through a long short-term memory network. The system dynamically decides whether to integrate sudden demand information based on whether the regional influence intensity exceeds a preset threshold. If it does, the sudden demand information is integrated to generate an enhanced trend vector. A convolutional neural network is used to extract spatial features from the enhanced trend vector. At the same time, based on the interaction between environmental influence and data quality, the dimension of the feature vector is dynamically adjusted to obtain optimized trend representation data. A weighted trend model is constructed by fusing the weight distributions of multiple factors through an attention mechanism. Based on the prediction results of the weighted trend model at different time scales, it is determined whether the prediction accuracy is lower than a preset threshold. If it is lower, the regional action parameters are adjusted back to determine the final prediction model. If the environmental impact fluctuates greatly, data quality filtering is automatically triggered to obtain stable load prediction results. Based on the stable load forecast results output by the final forecast model, a coordinated mapping between multiple regions is generated, and real-time verification is performed based on sudden changes in demand to output a comprehensive power load optimization scheme.

[0009] Preferably, if the environmental impact fluctuates significantly, data quality filtering is automatically triggered to obtain stable load forecast results, specifically including: The final prediction model is used to obtain a stable power supply simulation output. The model is then calibrated to address potential risks of resource waste and to determine whether environmental fluctuations are significant. If environmental fluctuations are significant, data quality filtering is automatically triggered to obtain a stable load prediction result.

[0010] Preferably, the step of extracting preliminary trend feature vectors of the time series based on historical load data and multi-source factor data through a long short-term memory network includes: Load-related records are retrieved from stored historical data, and the data is categorized and organized according to different time periods to obtain a structured load dataset. The load trends contained therein are analyzed, and statistical methods are used to identify the main change patterns and determine the key trend directions. Based on key trend directions and combined with data records from multiple sources, the dynamic correlation between the two is extracted to identify the main external variables affecting the trend. If the identified major external variables exceed the preset threshold range, the time series data is segmented to obtain the segmented sequence fragments, and a long short-term memory network is used for deep learning processing to generate the corresponding trend feature vectors. By combining the generated trend feature vector with the results of dynamic correlation, the load trend is further refined and decomposed to obtain multi-dimensional feature representation data. Then, the load changes of subsequent time series are predicted and simulated to obtain the predicted trend direction.

[0011] Preferably, the step of dynamically determining whether to fuse sudden demand information based on whether the regional effect intensity exceeds a preset threshold, and if it does, fusing the sudden demand information to generate an enhanced trend vector, includes: Obtain gridded load density data within the geographic range corresponding to the preliminary trend feature vector, and perform spatial aggregation calculation on the weighted average of the load values ​​of adjacent grids in the data to obtain load distribution data under regional action and its corresponding regional action intensity index. The regional action intensity index is compared with a preset threshold. If the regional action intensity index exceeds the preset threshold, the collection of sudden demand information within the synchronization time window is triggered. The sudden demand information comes from the external event report data stream. Principal component analysis is used to identify the demand pulse signal from the external event report data stream. The input of the principal component analysis is the numerical sequence of the event report data stream, and the output is the dimensionality-reduced demand pulse signal. The demand pulse signal is normalized, and the weight coefficient of the demand pulse signal in the fusion process is dynamically calculated based on the proportion of the regional effect intensity index exceeding the threshold and a fixed base. Through vector space mapping, the weighted demand pulse signal is converted into a demand influence vector with the same dimension as the preliminary trend feature vector, wherein the vector space mapping projects the signal to the corresponding dimension of the vector space. Within the vector space, the preliminary trend feature vector and the demand influence vector are superimposed. The superposition operation follows a linear weighting rule and is determined by adding elements one by one. The result is determined as the enhanced trend vector.

[0012] Preferably, the step of using a convolutional neural network to extract spatial features from the enhanced trend vector, and dynamically adjusting the feature vector dimension based on the interaction between environmental influence and data quality to obtain optimized trend representation data, includes: An enhanced trend vector is obtained, spatial sequence data is extracted from the trend vector, and a convolutional neural network is used to perform convolution operations on the spatial sequence data to obtain preliminary spatial feature representation data. For the preliminary spatial feature representation data, environmental impact factor data and data quality indicators are collected, and the two are multiplied to calculate the interaction coefficient. It is then determined whether the interaction coefficient exceeds a preset threshold to obtain the interaction judgment result. If the interaction judgment result indicates that adjustment is needed, then the preliminary spatial feature representation data is subjected to matrix multiplication dimension transformation according to the interaction coefficient to obtain the adjusted spatial feature representation data; The adjusted spatial feature representation data is used to perform vector weighted summation and fusion of the load distribution aggregate data and the sudden demand fusion information to calculate the grid load balance index and determine the dynamic demand calibration vector. The dynamic demand calibration vector is superimposed onto the adjusted spatial feature representation data to obtain the optimized trend representation data.

[0013] Preferably, the step of fusing the weight distributions of multiple factors through an attention mechanism to construct a weighted trend model includes: By obtaining historical samples of electricity load, sequence data is acquired to obtain initial trend representation data; An attention mechanism is used to fuse multiple sources of factors, extract features from the initial trend representation data, and determine the weight distribution. Based on the weight distribution and combined with historical samples, the trend representation data is optimized to obtain the weighted trend; For the weighted trend, the difference between the peak and trough values ​​is calculated from the load sequence. If the difference exceeds a preset threshold, the parameters of the multi-source factors are adjusted, and the stability of the model is judged by comparing the trend fluctuations before and after the adjustment. The stability of the model is obtained, and load peak-valley regulation is integrated to construct a weighted trend model.

[0014] Preferably, the step of iteratively updating the dynamic correlation based on the weighted trend model at different time scales, determining whether the prediction accuracy is lower than a preset threshold, and if so, backtracking to adjust the regional action parameters to determine the final prediction model includes: Obtain the prediction result sequence of the weighted trend model at three time scales: hourly, daily, and weekly, as well as the actual load observation sequence at the corresponding time scales; Based on the predicted result sequence and the observation sequence, calculate the prediction error sequence for each time scale; Using the Pearson correlation coefficient, the dynamic correlation between hourly scale error and daily scale error, and between daily scale error and weekly scale error, is analyzed to obtain a correlation strength matrix, which is used to describe the degree of mutual influence of prediction errors at different time scales. Based on the correlation strength matrix, the dynamic weight allocation mechanism in the weighted trend model is iteratively updated to adjust the model's dependence weight on historical data at different time scales; wherein, if the correlation strength between the hourly and daily scales is high, the weight of hourly historical data in daily scale prediction is increased. After iterative updates, the first updated model is obtained. Iterative optimization and convergence determination based on parameter backtracking mechanism are then performed to determine the final prediction model.

[0015] Preferably, after the iterative update, a first updated model is obtained, and iterative optimization and convergence determination based on a parameter backtracking mechanism are performed to determine the final prediction model, including: Using the first update model, load forecasting is performed for a preset future time period to generate a second forecast sequence; The second predicted sequence is compared with the new actual observed sequence to calculate the overall prediction accuracy; Determine whether the overall prediction accuracy is lower than a preset accuracy threshold; If the accuracy is lower than the threshold, the parameter backtracking mechanism is triggered; The parameter backtracking mechanism targets specific regions with large prediction errors. Specifically, it extracts the predicted load curve for the region from the second prediction sequence and the actual load curve for the corresponding region from the actual observation sequence. It analyzes the differences between the two curves, decomposes the error components by point-by-point comparison, and identifies the systematic deviation mainly caused by insufficient characterization of regional load features. Based on the direction and magnitude of the systematic deviation, it adjusts the action parameter vector corresponding to the region in the first update model in reverse. Using the adjusted action parameter vector, it updates the first update model to obtain the second update model. Repeat the above prediction, accuracy judgment and parameter backtracking process until the overall prediction accuracy of three consecutive iterations is not lower than the preset accuracy threshold, or the adjustment range of the model parameters is less than a preset convergence threshold. Once the model is determined to be stable, the model parameters obtained from the last iteration are used as the parameters for the final prediction model.

[0016] Preferably, the step of obtaining a stable power supply simulation output from the final prediction model, calibrating against potential risks of resource waste, determining whether environmental impact fluctuations are significant, and if so, incorporating data quality filtering to obtain a stable load prediction result includes: A stable analog output is obtained, and the simulation output is calibrated by comparing it with a preset risk threshold to address the potential risk of resource waste. For the calibrated simulated output, it is determined whether the environmental influence fluctuation is large. If the environmental influence fluctuation is large, data quality filtering is incorporated, which is achieved by removing outliers, and the filtered output data is determined. From the filtered output data, the distribution characteristics of the power grid load balancing are obtained and determined by calculating the load variance, thus determining the balanced load data. For the balanced load data, environmental fluctuation calibration is incorporated by adjusting the fluctuation coefficient to obtain the fluctuation calibration result; From the fluctuation calibration results, the stable load prediction results are obtained.

[0017] Preferably, the step of generating a coordinated mapping between multiple regions based on the stable load forecast results output by the final forecast model, and performing real-time verification based on changes in sudden demand, to output a comprehensive power load optimization scheme includes: Based on the stable load forecast results output by the final forecast model, a coordinated mapping between multiple regions is generated. A preliminary integration is performed on the preset regional energy sharing agreement, which includes inter-regional energy exchange rules, to obtain a set of mapping relationships. The mapping relationship set is obtained, and real-time verification is performed for sudden changes in demand. If the verification result exceeds the preset threshold, the mapping parameters are adjusted by adding or subtracting parameters to obtain the verified mapping. Demand fluctuation response data is extracted from the verified mapping, and balance calculation is performed using a preset regional interaction relationship. The calculation is determined by a weighted summation method based on the interaction relationship to determine the power balance adjustment value. Based on the power balance adjustment value, the preset scheme adjustment logic and stability evaluation index are integrated. The integration is carried out by matching the logical rule index to determine the resource allocation optimization direction and obtain the optimized intermediate scheme. By integrating the load forecasting results with the real-time verification process through the optimized intermediate scheme, the comprehensive optimization scheme is finally calibrated. The calibration is achieved through result comparison and adjustment, and the comprehensive power load optimization scheme is output.

[0018] The technical solutions provided by the embodiments of the present invention have the following beneficial effects: This invention extracts time-series trend features using a Long Short-Term Memory (LSTM) network, combines this with a convolutional neural network for spatial feature extraction, and introduces an attention mechanism to optimize the weight distribution of multi-source factors. Simultaneously, it dynamically adjusts the trend vector based on regional effects and sudden demand fluctuations, ensuring the prediction model adapts to changes at different time scales. If prediction accuracy or environmental fluctuations exceed thresholds, this invention achieves adaptive model calibration and stable output through backtracking parameter adjustments and data quality filtering, ultimately generating multi-regional coordinated mapping and comprehensive optimization schemes. This invention integrates multi-dimensional features with a dynamic iterative update mechanism to form a dynamic adaptive system, significantly improving the accuracy of load forecasting and power supply stability under the influence of multiple complex factors, and realizing the optimal allocation of power resources.

[0019] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0020] Figure 1 The flowchart shows the power load prediction method based on the fusion of fractional-order calculus control and graph neural network of the present invention. Figure 2 This is a schematic diagram of the power load forecasting method based on the fusion of fractional-order calculus control and graph neural network of the present invention; Figure 3 This is another schematic diagram of the power load prediction method based on the fusion of fractional-order calculus control and graph neural network of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figures 1-3 As shown in the figure, an embodiment of the present invention provides a power load forecasting method based on the fusion of fractional-order calculus control and graph neural networks, which may specifically include: Step S101: Based on historical load data and multi-source factor data, extract the preliminary trend feature vector of the time series through the long short-term memory network, and dynamically decide whether to integrate the sudden demand information according to whether the regional influence intensity exceeds the preset threshold. If it exceeds the threshold, integrate the sudden demand information to generate an enhanced trend vector.

[0023] include: (1) By extracting the dynamic correlation between load trends and multi-source factors from historical data, and using long short-term memory networks to process time-scale sequence data, a preliminary trend feature vector is obtained.

[0024] Specifically, load-related records are retrieved from stored historical data, and the data is categorized and organized according to different time periods to obtain a structured load dataset. For this structured load dataset, the load trends are analyzed, and statistical methods are used to identify the main change patterns and determine the key trend directions. Based on the key trend directions, and combined with data records from multiple sources, the dynamic correlation between these factors is extracted to identify the main external variables influencing the trend. If the identified main external variables exceed a preset threshold, the time series data is segmented to obtain segmented sequence fragments. For these segmented sequence fragments, a Long Short-Term Memory (LSTM) network is used for deep learning to generate corresponding trend feature vectors. Using these generated trend feature vectors, combined with the results of dynamic correlation, the load trend is further refined and decomposed to obtain multi-dimensional feature representation data. Based on this multi-dimensional feature representation data, load changes in subsequent time series are predicted and simulated to determine the predicted trend direction.

[0025] First, extract the dynamic correlation between load trends and multiple factors from historical data.

[0026] For example, hourly electricity load data for a certain area over the past 365 days, along with multi-source data such as temperature, humidity, holiday markers, and industrial production index for the same period, were collected. Pearson correlation coefficient analysis was used to calculate a correlation coefficient of 0.78 between the load and the highest temperature of the day, and a correlation coefficient of 0.65 between the load and the industrial production index, thus quantifying the dynamic correlation.

[0027] Next, a Long Short-Term Memory (LSTM) network is used to process these time-series data. For example, normalized data from 30 consecutive days (720 time points) is used as the input sequence. The LSTM network is configured with two hidden layers, each with 128 neurons, and trained using the Adam optimizer at a learning rate of 0.001. The network predicts the load trend for day 30 using multi-source data from the previous 29 days. The network learns long-term dependencies through a gating mechanism.

[0028] For example, the forget gate learns that when the holiday flag is 1, the load pattern of the previous workday needs to be weakened, and the weight may be adjusted to 0.3. After training, the output of the LSTM hidden state at the last time step is used as the initial trend feature vector. This is a 128-dimensional vector, which contains feature dimensions that, for example, represent periodic fluctuations and temperature sensitivity, with values ​​that may be [0.12, -0.45, ..., 0.78]. These vectors can then be used for downstream load forecasting or anomaly detection tasks, forming a complete automated process from data association analysis to deep feature extraction.

[0029] (2) Based on the preliminary trend feature vector, obtain the load distribution data under the regional effect. If the intensity of the regional effect exceeds the preset threshold, integrate the sudden demand information to determine the enhanced trend vector.

[0030] Specifically, gridded load density data within the geographical area corresponding to the preliminary trend feature vector is obtained. This gridded load density data includes the load value of each spatial grid over a time series. By spatially aggregating the weighted average of the load values ​​of adjacent grids in the gridded load density data, load distribution data under regional influence and its corresponding regional influence intensity index are obtained. The regional influence intensity index is compared with a preset influence intensity threshold. If the regional influence intensity index exceeds the preset threshold, the collection of sudden demand information within the synchronization time window is triggered. This sudden demand information originates from an external event report data stream. Principal component analysis (PCA) is used to identify demand pulse signals from the external event report data stream. The PCA input is the numerical sequence of the event report data stream, and the output is the dimensionality-reduced demand pulse signal. The demand pulse signal is normalized, and based on the proportion of the regional influence intensity index exceeding the threshold, the weight coefficient of the demand pulse signal in the fusion process is dynamically calculated by multiplying the proportion by a fixed base. Through vector space mapping, the weighted demand pulse signal is converted into a demand influence vector with the same dimension as the preliminary trend feature vector. Vector space mapping projects the signal onto the corresponding dimension of the vector space. Within the vector space, the preliminary trend feature vector and the demand influence vector are superimposed. The superposition operation follows a linear weighting rule and is determined by adding elements one by one. The result is determined as the enhanced trend vector.

[0031] In one implementation, a preliminary trend feature vector is first obtained. This vector is a numerical representation obtained after analyzing historical load data and is used to describe the overall direction and magnitude of load changes.

[0032] For example, in a power load management system, a preliminary trend feature vector can be derived through statistical calculations of electricity consumption data from the past 24 hours. This vector includes components such as average load value and peak fluctuation rate, and is represented in vector form as [average value, fluctuation rate]. This helps to capture potential load patterns in subsequent steps. Further, based on the preliminary trend feature vector, load distribution data under regional effects is obtained. Regional effects refer to the impact of specific geographical or environmental factors on the load, such as fluctuations in electricity demand caused by weather changes or regional events.

[0033] Specifically, in energy distribution systems, load distribution data is collected from multiple nodes within a monitoring area. For example, meteorological data is combined to calculate the impact of wind or temperature on electrical equipment, thereby generating a distribution map that displays load density values ​​for different sub-regions. This data acquisition process ensures the correlation between the vector and the actual environment, improving the accuracy of predictions.

[0034] It should be noted that the regional effect intensity is an indicator obtained by quantifying these influencing factors. For example, parameters such as wind speed and temperature deviation are weighted and summed to form an intensity score.

[0035] In one possible implementation, if the intensity score exceeds a preset threshold, such as 80 (based on historical experience), but not limited to this, it indicates that regional factors have a significant impact on the load, and further processing is required to avoid prediction bias.

[0036] Preferably, if the intensity of the regional impact exceeds a preset threshold, then sudden demand information is fused. Sudden demand information refers to the additional load demand caused by a sudden event, such as power outage restoration or a surge in electricity consumption due to a large-scale event. The specific fusion process includes converting the sudden information into a supplementary vector. For example, an impact coefficient is calculated based on the event type and expected duration, and then weighted and merged with the preliminary trend feature vector to form an enhanced trend vector. This fusion can employ a linear combination method to ensure that the enhanced vector more comprehensively reflects actual demand.

[0037] For example, in an urban power grid, assume the initial trend feature vector is [50, 0.2], representing an average load of 50 units and a volatility of 0.2. If the regional impact intensity exceeds a threshold due to heavy rainfall, then sudden demand information such as emergency lighting demand is fused to calculate an enhancement vector [65, 0.3], thereby optimizing resource scheduling.

[0038] In one embodiment, the implementation of this process can be extended to energy networks of different sizes.

[0039] For example, in small community power grids, the initial vector is acquired based on local data, and the regional impact strength threshold is set to a low value for sensitive response; while in large urban power grids, the threshold is higher, and more types of emergency information, such as traffic incident data, are incorporated to adapt to complex environments.

[0040] Specifically, enhancing the determination of the trend vector helps improve the accuracy of subsequent load forecasting. Through this mechanism, the system can dynamically adjust the vector, reducing errors caused by regional and sudden factors, and achieving more stable supply allocation in energy management. Furthermore, in the fusion step, the source of sudden demand information can be real-time sensors or external databases, such as data provided by a weather warning system, ensuring the timeliness of the information. In practical applications, this step is executed through a software module, taking the initial vector and intensity judgment as input, and outputting the enhanced vector.

[0041] For example, when dealing with peak loads, if the intensity does not exceed a threshold, the initial vector is used directly; otherwise, the fused vector can be used to generate adjustment strategies, such as increasing the investment of backup power, thereby improving the system's responsiveness in terms of business operations.

[0042] Understandably, the versatility of this technical solution is reflected in various energy scenarios, such as load management in industrial or residential areas. It can adapt to specific thresholds and fusion rules, ensuring the flexibility and effectiveness of the vector enhancement process.

[0043] Step S102: Spatial features are extracted from the enhanced trend vector using a convolutional neural network. Simultaneously, the dimension of the feature vector is dynamically adjusted based on the interaction between environmental influence and data quality to obtain optimized trend representation data.

[0044] An enhanced trend vector is obtained, and spatial sequence data is extracted from the trend vector. A convolutional neural network is used to perform convolution operations on the spatial sequence data to obtain preliminary spatial feature representation data. For the preliminary spatial feature representation data, environmental impact factor data and data quality indicators are collected. An interaction coefficient is calculated by multiplying the environmental impact factor data and the data quality indicators. It is determined whether the interaction coefficient exceeds a preset threshold to obtain an interaction judgment result. If the interaction judgment result indicates that adjustment is needed, a matrix multiplication dimensionality transformation is performed on the preliminary spatial feature representation data according to the interaction coefficient to obtain adjusted spatial feature representation data. The adjusted spatial feature representation data is then used to perform vector weighted summation and fusion of the load distribution aggregation data and the sudden demand fusion information to calculate the grid load balancing index and determine the dynamic demand calibration vector. The dynamic demand calibration vector is superimposed on the adjusted spatial feature representation data to obtain optimized trend representation data.

[0045] In one implementation, a convolutional neural network is used to extract spatial features from the enhanced trend vectors. This process first takes the enhanced trend vectors as input data; these vectors typically represent the temporal variations and regional impacts of power load. In a power load management system, the convolutional neural network captures the spatial correlation of vectors through multiple layers of convolutional operations. For example, the network structure includes an input layer, multiple convolutional layers, and pooling layers, where the convolutional layers use one-dimensional convolutional kernels to slide and scan vector components to extract local patterns such as continuous fluctuations in load peaks, but is not limited to this.

[0046] Specifically, the convolution operation calculates a weighted sum of each component with its neighboring components to form a feature map, which is used to identify the implicit spatial structure of the trend vector. This extraction helps reveal the potential patterns of load distribution in energy distribution systems. For example, in urban power grids, the extracted features can highlight the spatial clustering of peak loads, but it is not limited to this. In this way, the system can obtain richer spatial information from the enhanced vector, providing a foundation for subsequent processing. Furthermore, considering the interaction between environmental impact and data quality, it is determined whether the vector dimension needs to be adjusted. Environmental impact refers to the external effects of factors such as temperature or events on the load, while data quality involves the integrity and noise level of the collected data.

[0047] In one embodiment of the invention, the interaction effect is first quantified, for example by calculating the ratio of the weight of environmental factors to data noise to form an interaction index, but not limited thereto. If the index shows that environmental influences amplify data instability, the dimensions need to be adjusted. The specific process includes evaluating the contribution of each dimension of the vector. If some dimensions are excessively affected by interaction interference, such as bias caused by missing data, the dimensions are reduced using principal component analysis to ensure a more refined vector.

[0048] It should be noted that this judgment is based on preset rules, such as triggering adjustments when interaction indicators exceed a certain value. In small community power grids, this process can quickly respond to changes in the local environment, but it is not limited to this.

[0049] Preferably, after determining that adjustment is needed, dimensional adjustment is performed to obtain optimized trend representation data.

[0050] For example, in a large urban power grid, if an interactive assessment confirms that an adjustment is needed, the vector is reduced from high-dimensional to low-dimensional, retaining core components such as average load and volatility, thus forming an optimized representation of the data. This representation is more suitable for subsequent load forecasting and avoids interference from redundant information.

[0051] In one embodiment, the optimization trend representation data can be directly used in the resource scheduling module, demonstrating the applicability of the technical solution in energy management.

[0052] Specifically, the spatial feature extraction of convolutional neural networks demonstrates versatility across different energy scenarios. In industrial load management, the network can process vectors containing sudden demand and extract spatial features to identify the distribution patterns of production peaks; while in residential power grids, it focuses on daily fluctuations and extracts features to optimize supply allocation.

[0053] Understandably, this extraction process ensures consistency through standardized input, preventing data scale differences from affecting the results. Furthermore, the mechanism for judging the interaction between environmental impact and data quality can be extended to real-time monitoring systems.

[0054] For example, by integrating sensor data, the system calculates interaction values. If environmental factors, such as heavy rain, cause a decline in data quality, the vector dimension is adjusted to 80% of the original dimension to obtain a more stable and optimized representation. In one implementation, this adjusted representation can improve the robustness of the predictive model, enabling more reliable energy allocation in business operations.

[0055] For example, when dealing with peak loads, if the interactive judgment requires no adjustment, the extracted features are used directly to obtain the optimized trend representation data; otherwise, the adjusted representation data integrates environmental adaptability, ensuring an accurate description of load trends under complex conditions. The application of this method in energy networks demonstrates the flexibility of the technical solution.

[0056] In one embodiment, the entire process begins with an augmented vector input, proceeds through convolutional extraction and interactive judgment, and finally outputs optimized trend representation data for downstream tasks such as demand forecasting.

[0057] Specifically, the optimized representation of data can be in the form of a low-dimensional vector containing refined load characteristics, supporting dynamic adjustment strategies in the power system.

[0058] Step S103: The weighted trend model is constructed by fusing the weight distribution of multiple factors through an attention mechanism.

[0059] Specifically, by optimizing trend representation data and combining it with historical samples of electricity load, an attention mechanism is used to focus on the weight distribution of multiple factors to obtain a weighted trend model.

[0060] Specifically, historical samples of electricity load are used to obtain sequential data, resulting in initial trend representation data. An attention mechanism is employed to focus on multiple factors, extracting features from the initial trend representation data and determining the weight distribution. Based on this weight distribution and combined with historical samples, the trend representation data is optimized to obtain a weighted trend. For this weighted trend, the difference between peak and trough values ​​in the load sequence is calculated. If the difference exceeds a preset threshold, the parameters of the multiple factors are adjusted, and the model stability is determined by comparing the trend fluctuations before and after the adjustment. The model stability is then obtained, and load peak-valley regulation is integrated to obtain a weighted trend model.

[0061] In one implementation, the power load forecasting system first collects historical sample data derived from power grid operation records, including load values, timestamps, and relevant environmental variables over a past period. This ensures the integrity of the underlying data for trend representation.

[0062] Specifically, historical samples can be extracted from the database of the power dispatch center, such as selecting hourly load data from the past year as input, but are not limited to this. This collection process helps to capture the seasonal and cyclical changes in load, providing support for subsequent optimization. Furthermore, optimizing trend representation data involves preprocessing and mathematical transformations of historical samples.

[0063] For example, the original load sequence is decomposed into trend components, seasonal components, and residual components, using moving average or wavelet transform methods. In the field of power load, this optimization can be applied to urban power grid scenarios to help identify peak-valley load patterns.

[0064] It should be noted that optimizing the trend representation data aims to reduce noise interference, thereby improving the robustness of the model. Through this step, the system can generate smoother trend sequences, laying the foundation for incorporating multi-source factors.

[0065] In one possible implementation, an attention mechanism is introduced to focus on the weight distribution of multi-source factors. These factors include weather conditions, holiday influences, and economic activity levels, all of which significantly impact electricity load. The principle of the attention mechanism is to assign weights by computing query, key, and value vectors. Specifically, the multi-source factors are first converted into embedding vectors, and then attention scores are calculated—that is, a softmax function is applied to dot product similarity. This mechanism allows the model to dynamically emphasize important factors, such as increasing the weight of weather factors during hot weather, but not limited to this. In the operational process of electricity load forecasting, this means the system can adjust its focus based on real-time input, avoiding interference from irrelevant factors, thus obtaining an accurate weight distribution. The detailed implementation of this process involves multiple attention heads computing in parallel to capture dependencies across different dimensions; for example, one head might focus on time factors, while another focuses on external events, but not limited to this. Through this parallel structure, the model can handle complex multi-source interactions, improving forecast accuracy. In practical applications, such as in industrial power grids, when economic activity levels rise, the attention mechanism automatically increases the weight of that factor, ensuring the trend model reflects actual load changes, but not limited to this. The effect of this technique of focusing on weight distribution is that it improves the model's ability to adapt to uncertainty, especially in scenarios with large load fluctuations.

[0066] Preferably, a weighted trend model is constructed by combining historical samples of electricity load with the results of the attention mechanism.

[0067] Specifically, the optimized trend representation data is used as the base sequence, and then the weights derived from attention are applied to the multi-source factors for weighted summation to form a weighted input.

[0068] For example, in one embodiment, if the historical sample sequence is represented as S(t) and the weights of the multi-source factors are W_i, then the weighted trend model can be expressed as M(t) = S(t) + Σ W_i * F_i(t), where F_i(t) is the value of each factor at time t. This construction process is applicable to residential scenarios in power load forecasting, such as integrating historical peak loads with weather weights to predict summer air conditioning loads. In this way, the model can generate continuous trend forecast curves, supporting grid dispatching decisions, but is not limited to this.

[0069] Understandably, in another embodiment, the system is extended to a regional power grid scenario, where historical samples include load data from multiple sub-regions. The attention mechanism here focuses on cross-regional factors, such as transmission line load and neighboring area influences. The weight distribution is calculated through a self-attention layer, ensuring the model captures spatial dependencies. The detailed process involves first normalizing the samples, then inputting them into the attention module, and finally applying the output weight matrix to trend optimization. This implementation demonstrates the versatility of the technology, applicable to power systems of varying sizes. Furthermore, for the weight distribution of multi-source factors, the system can introduce residual connections to enhance stability.

[0070] For example, adding a residual path after attention calculation avoids the vanishing gradient problem. In the business process of electricity load, this helps with long-term forecasting, such as annual load trend planning.

[0071] Specifically, the weight distribution process involves iteratively updating the weights until they converge to a stable distribution, thus obtaining a reliable weighted model.

[0072] In one embodiment, the weighted trend model is trained using historical samples as the training set and mean squared error as the loss function. The model learns attention weights progressively by optimizing parameters through gradient descent. This is applicable to power systems with wind power integration.

[0073] For example, when multiple factors include wind speed, the attention mechanism dynamically adjusts its weights, and the model output reflects the impact of renewable energy on the load. A key detail of this training process is batch processing of samples to ensure computational efficiency.

[0074] For example, another embodiment focuses on emergency load forecasting scenarios, such as during natural disasters. Optimized trend representation data is combined with short-term historical samples, and an attention mechanism prioritizes disaster factors, such as storm intensity, to generate a weighted model for rapid response. This approach demonstrates the applicability of the technology under high-pressure situations.

[0075] It should be noted that the overall methodology ensures consistency in its logical flow from data acquisition to model output. Based on the above steps, the model's effectiveness is further validated, for example, by calculating the prediction accuracy on a test dataset, demonstrating improvement under multi-source interference, but not limited to this.

[0076] In one possible implementation, the system can be integrated into a smart grid platform, with the weighted trend model output used for real-time scheduling. This integration allows the technology to demonstrate flexibility in various scenarios within the power sector, such as transitions from urban to rural power grids, but not limited to these applications.

[0077] Step S104: Based on the prediction results of the weighted trend model at different time scales, determine whether the prediction accuracy is lower than a preset threshold. If it is lower, backtrack and adjust the regional action parameters to determine the final prediction model. If the environmental impact fluctuates greatly, automatically trigger data quality filtering to obtain stable load prediction results.

[0078] include: (1) Based on the weighted trend model, the dynamic correlation under different time scales is iteratively updated. If the prediction accuracy is lower than the preset threshold, the regional action parameters are adjusted back to determine the final prediction model.

[0079] Specifically, the prediction result sequences of the weighted trend model at three time scales—hourly, daily, and weekly—are obtained, along with the actual load observation sequences at the corresponding time scales. Based on the prediction result sequences and observation sequences, the prediction error sequences for each time scale are calculated. Using the Pearson correlation coefficient, the dynamic correlation between hourly-scale errors and daily-scale errors, and between daily-scale errors and weekly-scale errors, is analyzed to obtain a correlation strength matrix. This correlation strength matrix describes the degree of mutual influence between prediction errors at different time scales. Based on the correlation strength matrix, the dynamic weight allocation mechanism in the weighted trend model is iteratively updated. The update process adjusts the model's dependence weights on historical data at different time scales. If the correlation strength between the hourly and daily scales is high, the weight of hourly-scale historical data in the daily-scale prediction is increased. After iterative updates, a first updated model is obtained. Iterative optimization and convergence determination based on a parameter backtracking mechanism are performed to determine the final prediction model. That is, using the first updated model, load prediction is performed for a preset future period, generating a second prediction sequence. The second prediction sequence is compared with the new actual observation sequence to calculate the overall prediction accuracy. It is then determined whether the overall prediction accuracy is lower than a preset accuracy threshold. If the accuracy is below a threshold, a parameter backtracking mechanism is triggered. This mechanism targets specific regions with large prediction errors. The predicted load curve for the region is extracted from the second prediction sequence, and the actual load curve for the corresponding region is extracted from the actual observation sequence. The differences between the two curves are analyzed, and error components are decomposed through point-by-point comparison to identify systematic deviations primarily caused by insufficient characterization of regional load features. Based on the direction and magnitude of these systematic deviations, the action parameter vector corresponding to that region in the first update model is adjusted in reverse. The adjusted action parameter vector is used to update the first update model, resulting in a second update model. This prediction and accuracy judgment process, along with the parameter backtracking process, is continuously executed until the overall prediction accuracy for three consecutive iterations is not lower than the preset accuracy threshold, or the model parameter adjustment magnitude is less than a preset convergence threshold. At this point, the model state is considered stable, and the model parameters obtained from the last iteration are determined as the parameters of the final prediction model.

[0080] In one implementation, the power load forecasting system is based on a weighted trend model and iteratively updates the dynamic correlations at different time scales.

[0081] Specifically, the weighted trend model is a sequence model constructed from historical samples and attention weights to capture load trends. Dynamic correlations refer to the temporal interactions between load values ​​and multi-source factors such as weather or economic activity; for example, short-term correlations between load and temperature at the hourly scale, and long-term correlations with weekday patterns at the daily scale, but not limited to these. The iterative update process calculates the strength of these correlations iteratively, first inputting the current model output, then comparing the correlation vectors at different scales, and adjusting model parameters to enhance consistency. When applied in urban power grids, this approach can handle rapid changes in peak load and ensure gradual model convergence. Furthermore, the system defines scales as hourly, daily, or weekly levels for different time scales.

[0082] For example, on an hourly scale, dynamic associations are quantified by calculating the autocorrelation function of the load series, such as by decomposing the series into short-term fluctuation components and using gradient methods to gradually adjust the association weights during updates, but not limited to this.

[0083] In one possible implementation, for a weekly scale, the system considers the dynamic impact of seasonal factors, such as the correlation between winter heating load and temperature, and updates by accumulating historical data deviations during iterations, but is not limited to this. This scale distinction helps the model adapt to the periodicity of the load, ensuring that the update process covers a comprehensive correlation from short-term fluctuations to long-term trends. In the business of electricity load forecasting, this means that the system can dynamically respond to sudden events, such as load changes during holidays, but is not limited to this.

[0084] Preferably, if the prediction accuracy is lower than a preset threshold, the system initiates a step of retrospectively adjusting the regional impact parameters. Prediction accuracy is calculated using mean squared error or absolute percentage error and compared to a threshold such as 0.05, but is not limited to these methods. Regional impact parameters refer to the influence coefficients of a specific geographical or sub-grid region on the overall load. For example, parameters for industrial areas represent the weighted contribution of economic activity in that region to the total load, but are not limited to these methods. Retrospective adjustment involves a reverse tracing iterative step to identify parameter values ​​leading to low accuracy, and then correcting them using a weighted averaging method, such as reducing the parameter weights of high-biased regions, but is not limited to these methods. This mechanism, applied in regional power grid scenarios, ensures the model's sensitivity to cross-regional impacts.

[0085] Specifically, the backtracking adjustment process includes several stages. First, the iteration history is recorded. Then, if the accuracy is lower than the threshold, the process is backtracked to the previous iteration point to analyze the deviation distribution of regional parameters.

[0086] It should be noted that the calculation of regional action parameters is based on a spatial model. For example, the power grid is divided into sub-regions, and the parameters of each sub-region are determined by the historical load ratio. If the parameters of the industrial sub-region cause an increase in error during backtracking, its weight ratio is reduced to 80% of the original value, but this is not the only possibility.

[0087] In one embodiment, this adjustment is applied to wind-intensive areas, with parameters reflecting the regional impact of wind speed on load. Backtracking improves model accuracy and supports emergency dispatching. This detailed process demonstrates the logic behind parameter adjustment, ensuring model stability.

[0088] Understandably, when determining the final prediction model, the system integrates the results of iterative updates and backtracking adjustments to form a stable output model.

[0089] For example, in one embodiment, after multiple iterations, if the accuracy at all scales is above a threshold, the current parameters are fixed, and a final model is generated for future load forecasting. In a residential power grid scenario, this means the model can predict peak air conditioning loads, avoiding overfitting based on adjusted regional parameters. Furthermore, the system can introduce a validation set to test the final model, for example, using data from the past month to verify the accuracy of capturing dynamic correlations, but is not limited to this. This determination process emphasizes the model's robustness, maintaining high accuracy under multi-source interference. In another implementation, for load forecasting of industrial power grids, iterative updates focus on the time-scale dynamic correlations of economic activities, such as the difference between weekdays and weekends. The system first calculates the correlation matrix, then iteratively updates the weights; if the accuracy is low, it backtracks to adjust regional parameters, for example, correcting load contribution parameters for densely populated factory areas, but is not limited to this. In this way, the model adapts to changes in the production cycle, ensuring that the forecast supports energy allocation.

[0090] For example, when the economy declines, the correlation weakens, the iterative process automatically reduces the weight of relevant parameters, and the backtracking step refines the parameters to match the actual downward trend. This application scenario highlights the flexibility of the technology. Furthermore, the detailed mechanism of backtracking adjustments involves parameter sensitivity analysis, for example calculating the derivative contribution of each regional parameter to the overall error, and prioritizing adjustments for parameters with larger contributions. In the business processes of power load, this helps handle uncertainties such as the impact of sudden power outages on a region.

[0091] Preferably, in one embodiment, the system is applied to a rural power grid, dynamically correlated with seasonal scales including agricultural irrigation, and adjusting rural area parameters during backtracking to reflect water resource factors, ensuring that the final model accurately predicts peak irrigation loads. This analysis process improves the model's adaptability to remote areas by progressively refining the parameters.

[0092] It should be noted that the determination of the final prediction model can be extended to include feedback loops, such as adding real-time data calibration after iteration, but is not limited to this.

[0093] In one possible implementation, for transmission network scenarios, model updates are dynamically correlated with temporal changes in transmission line load. If the accuracy is insufficient, cross-regional parameters are adjusted retrospectively, and the final output is used for grid balancing. This extension demonstrates the method's versatility in large-scale systems, enabling optimized prediction of load distribution.

[0094] In one embodiment, the system processes dynamic correlations across mixed time scales by updating hourly and daily correlations in parallel iterations and uniformly adjusting regional parameters during backtracking.

[0095] For example, in a commercial area power grid, the correlation includes traffic flow factors, and the adjusted model is determined as the final version, supporting peak-valley regulation. This parallel structure improves efficiency and ensures coordination of correlations at different scales.

[0096] Understandably, the entire process, from the initial weighted trend model to the final determination, forms a closed optimization chain. This applies to power systems with wind power integration.

[0097] For example, dynamic correlation involves wind speed fluctuations. If the accuracy is low after iterative updates, the wind field area parameters are adjusted retrospectively, and the final model is used to integrate renewable energy forecasts. This approach enables precise management of fluctuating loads in business operations, improving the grid's responsiveness.

[0098] (2) Obtain the simulated output of stable power supply from the final prediction model, calibrate against the potential risk of resource waste, and if the environmental impact fluctuates greatly, incorporate data quality filtering to obtain stable load prediction results.

[0099] Specifically, a stable power supply simulation output is obtained from the final prediction model. The model is then calibrated by comparing the simulation output with a preset risk threshold to identify potential resource waste risks, resulting in a calibrated simulation output. For this calibrated simulation output, if environmental fluctuations are significant, data quality filtering is implemented to remove outliers, determining the filtered output data. From the filtered output data, the distribution characteristics of the power grid load balance are obtained and determined by calculating the load variance, thus defining the balanced load data. For this balanced load data, environmental fluctuation calibration is implemented by adjusting the fluctuation coefficient, resulting in a fluctuation calibration result. From the fluctuation calibration result, a stable load prediction result is obtained.

[0100] In one embodiment, a specific implementation method for power system load forecasting is provided. This implementation method aims to obtain more stable and reliable load forecast results from the forecasting model through calibration and filtering mechanisms, so as to support the formulation of generation plans and grid dispatching, and reduce the waste of generation resources or the risk to power supply reliability caused by forecasting deviations.

[0101] Specifically, obtaining a stable power supply simulation output from the final prediction model refers to using a trained load forecasting model, inputting historical load data, date type, and preliminary environmental factor sequences to obtain a sequence of predicted load values ​​for a specific future period (e.g., the next 24 hours, but not limited to this). This predicted value sequence is the initial simulation output, and its stability has not yet been evaluated and optimized. Calibration to address the potential risks of resource waste is one of the core steps. In power systems, resource waste risks mainly manifest in two aspects: first, a sustained increase in predicted load compared to actual demand leads to generator units operating under uneconomical conditions for extended periods or resulting in wind and solar curtailment; second, a sustained decrease in predicted load compared to actual demand leads to insufficient system reserve capacity, potentially triggering power curtailment risks during sudden load surges, which is essentially a waste of invested resources. The calibration process first needs to quantify this risk.

[0102] For example, a risk indicator can be defined as the cumulative deviation between the predicted load and the actual load (which can be approximated by recent actual values), with risk thresholds set separately for positive deviations (over-prediction) and negative deviations (under-prediction). Calibration is triggered when the deviation exceeds the threshold for multiple consecutive prediction periods. One implementation of calibration is to introduce a feedback adjustment coefficient, which is dynamically calculated based on the statistical characteristics of historical prediction errors (such as moving average errors, but not limited to these). Multiplying the initial prediction sequence by this adjustment coefficient provides a gentle correction to the prediction trend, making it closer to the changing trajectory of actual demand, thereby reducing the risk of resource misallocation.

[0103] It should be noted that fluctuations in environmental impacts have a significant impact on the quality of load forecast data. Environmental impacts here mainly refer to external factors such as meteorological conditions (e.g., temperature, humidity, wind speed, but not limited to these), holiday effects, and sudden public events. Large fluctuations in environmental impacts can be determined by calculating the variance or range of key environmental factors (e.g., daily maximum temperature, but not limited to these) within a recent time window.

[0104] For example, if the temperature range over three consecutive days exceeds a preset threshold (e.g., 10 degrees Celsius, but not limited to this), it is considered a large temperature fluctuation. Furthermore, factors such as the proximity of holidays can be considered in a comprehensive assessment. If the environmental impact is determined to be large, data quality filtering is incorporated. This is because when the environment fluctuates drastically, the collected raw load data or related feature data may contain more noise and outliers. Data quality filtering involves multiple steps.

[0105] For example, outlier detection is performed first using statistical methods, such as calculating the moving average and standard deviation of the load data series. Data points deviating from the average by more than three standard deviations are marked as outliers. Instead of simply deleting marked outliers, they are repaired using linear interpolation or seasonal trend decomposition methods based on the data trends of the preceding and following periods. Secondly, for data gaps caused by communication interruptions, time series forecasting methods (such as the ARIMA model) are used to fill in the missing segments, ensuring data continuity. The filtered data serves as cleaner input, either re-involved in or used to assist in the calculation of the load forecasting model, or to correct parameters in the aforementioned calibration steps. Based on the above calibration and conditional data quality filtering process, the final load forecast results exhibit higher stability. This stability is reflected in improved forecast curve smoothness, reduced sensitivity to sudden noise, and better tracking of the true load change trend, avoiding large fluctuations caused by environmental interference or inherent model biases.

[0106] In one embodiment, for short-term load forecasting scenarios, this implementation helps to formulate more accurate intraday unit start-up and shutdown plans; for medium- and long-term load forecasting, it can provide a more reliable basis for power generation equipment maintenance plans and fuel procurement.

[0107] Preferably, the calibration threshold for the risk of resource waste can be set differently according to different seasons or different load types (such as residential electricity consumption and industrial electricity consumption).

[0108] For example, air conditioning loads are greatly affected by temperature in summer and have a low tolerance for negative deviations (under-prediction). Therefore, a stricter negative deviation threshold can be set, and once it is reached, a calibration mechanism can be quickly activated to correct the predicted value upwards to ensure power supply reliability. In another possible implementation, domain knowledge can be incorporated into the data quality filtering process.

[0109] For example, when a sharp drop in load data is detected during major holidays (such as the Spring Festival, but not limited to this) and is misjudged as abnormal by the model, a holiday flag can be introduced to inform the filtering algorithm that the load pattern during this period is normal, thereby avoiding erroneous corrections. This further improves the intelligence and accuracy of filtering. Through the organic combination of the above steps, this implementation constructs an adaptive load forecasting post-processing workflow. It not only relies on the initial forecasting model but also dynamically adjusts the forecast output by monitoring risks and environmental fluctuations in real time, and ensures the reliability of the data used. Thus, it continuously outputs stable load forecasting results in complex actual operating environments, providing solid data support for the economical and safe operation of the power system.

[0110] Step S105: Based on the stable load forecast results output by the final forecast model, generate a coordinated mapping between multiple regions, and perform real-time verification based on sudden demand changes to output a comprehensive power load optimization scheme.

[0111] Specifically, based on stable load forecasting results, a coordinated mapping between multiple regions is generated. Preliminary integration is performed based on a preset regional energy sharing agreement, which includes inter-regional energy exchange rules, resulting in a set of mapping relationships. This mapping relationship set is then acquired, and real-time verification is performed for sudden changes in demand. If the verification result exceeds a preset threshold, the mapping parameters are adjusted by adding or subtracting parameters, resulting in a post-verified mapping. Demand fluctuation response data is extracted from the post-verified mapping, and balance calculations are performed using preset regional interaction relationships. These calculations are determined through a weighted summation of interaction relationships, resulting in a power balance adjustment value. Based on this power balance adjustment value, preset scheme adjustment logic and stability evaluation indicators are integrated. This integration is performed using logical rule matching indicators to determine the direction of resource allocation optimization, resulting in an optimized intermediate scheme. Using this optimized intermediate scheme, load forecasting results and the real-time verification process are integrated, and a final calibration is performed on the comprehensive optimization scheme. This calibration is achieved through result comparison and adjustment, resulting in a comprehensive power load optimization scheme.

[0112] In one implementation, a multi-regional integrated load optimization scheme for a power system is provided. This implementation is applied to provincial or inter-provincial power dispatch centers, aiming to improve the overall power grid operation economy and security by coordinating the load characteristics of different power supply areas.

[0113] Specifically, the first step is to obtain power load forecast data for each region. This forecast data may come from historical load data combined with the output of a forecasting model that incorporates factors such as weather and holidays. Since a single forecasting model may fluctuate, multiple forecasting methods can be fused together to obtain stable load forecast results.

[0114] For example, time series analysis and neural network prediction are used simultaneously to predict the load of the same region for a specific future time period (such as the next hour), resulting in multiple predicted values. These predicted values ​​are then weighted and averaged, with the weights dynamically adjusted based on the recent prediction accuracy of each model. This fusion method smooths out the random errors of individual models, yielding a more stable and reliable regional load prediction benchmark. This step provides a reliable data foundation for subsequent multi-regional coordination. Based on the stable load prediction results obtained for each region, further coordination mapping between multiple regions needs to be generated. The essence of coordination mapping is to establish dynamic correlation rules between regional loads, describing how other regions can respond to maintain overall balance when the load in one region changes. One possible implementation is to calculate the electrical coupling strength and power support potential between regions based on the power grid topology and line transmission capacity limitations.

[0115] For example, a regional correlation matrix can be constructed, where each element represents the ratio of the maximum transferable load capacity between two regions after considering network constraints. Simultaneously, the load characteristic curves of each region need to be analyzed to identify regional pairs with complementary load peak and trough periods.

[0116] For example, the peak load in region A typically occurs during the day, while the peak load in region B occurs at night, thus exhibiting good intraday complementarity. Considering both network transmission capacity and load characteristic complementarity, the coordinated mapping can be concretized into a series of rules, such as "when region X experiences a predicted load surge, priority should be given to drawing power from region Y, which has a load margin and a strong transmission connection with it; secondly, consideration should be given to activating interruptible loads in region Z for regulation." After generating this mapping relationship, a contingency plan library for multi-regional load coordination is formed.

[0117] It should be noted that unexpected and sudden load changes often occur in power grid operation, such as temporary start-ups and shutdowns by large industrial users or surges in electricity consumption caused by extreme weather. Therefore, the effectiveness of the coordinated mapping must be verified in real time. The verification process is triggered when a significant deviation is detected between the real-time load of a certain area and the stable forecast value. The verification process is not a simple check, but rather a rapid simulation based on the actual operating state of the current power grid to determine whether the pre-set response strategies in the coordinated mapping are feasible.

[0118] For example, the system acquires real-time data on actual power generation output, line power flow, and reserve capacity for each region. Then, it overlays sudden demand onto the corresponding region and, based on the scheme specified in the coordination map, attempts to calculate and adjust the power generation plans of other regions or perform load shifting. This calculation process typically employs rapid power flow calculations or sensitivity analysis to determine within seconds whether the scheme will trigger grid overruns (such as line overloads or voltage overflows, but not limited to these). If the verification passes, it indicates that the sudden demand is manageable under the existing coordination framework; if the verification fails, the coordination plan is marked as invalid at the current moment, and the backup analysis process is initiated. Furthermore, after completing real-time verification of sudden demand changes, the system integrates all information to generate the final power load optimization scheme. This scheme not only includes the basic power generation plans for each region based on stability forecasts but also incorporates verified or adjusted inter-regional coordination support instructions.

[0119] For example, the optimization scheme explicitly provides the output recommendations for each power plant, the planned power flow values ​​for inter-regional interconnections, and a list of contingency measures for potential load fluctuations over several future scheduling periods. The scheme can be generated based on an optimization model that aims to minimize total generation cost or network losses, while simultaneously satisfying power balance constraints, unit operation constraints, and inter-regional coordination constraints that have been verified and confirmed to be feasible in real time. The resulting scheme is both forward-looking (based on stability predictions) and robust (verified under contingency conditions), achieving comprehensive optimization of multi-regional loads. In another embodiment, the generation of the coordination map can be more refined. In addition to considering physical grid constraints, economic signals can also be introduced.

[0120] For example, a cost-optimal coordination map is constructed based on the differences in real-time electricity prices or marginal generation costs across regions. When a region experiences a load increase, the coordination map prioritizes supporting neighboring regions with lower generation costs, thereby guiding economically efficient power flow. The real-time verification process also needs to take economic cost constraints into account to ensure that the adjustment plan is economically acceptable.

[0121] Preferably, to improve the speed of real-time verification, extensive offline simulation calculations can be performed in advance for common sudden demand scenarios and network operation modes, forming a lookup table of "safety domain" or "feasibility domain". During online verification, the system only needs to determine whether the current sudden demand size and the grid state point are within the feasible domain to quickly obtain the verification conclusion. This greatly shortens the online calculation time and meets the stringent real-time requirements of power dispatch. Furthermore, the coordination mapping is not static. It can be a dynamic file that is updated periodically (e.g., daily or weekly, but not limited to). Dispatchers can adjust and update the inter-regional association rules and support priorities based on the latest grid infrastructure commissioning status, unit maintenance plans, or seasonal load changes, ensuring that the coordination strategy always closely reflects the actual grid situation. Ultimately, through the implementation of the above steps, the power dispatch center can obtain a coordinated, reliable, and responsive comprehensive load optimization scheme. This scheme helps to smooth inter-regional load fluctuations, reduce the expensive reserve capacity reserved to cope with uncertainties, and thus improve the overall operational economic efficiency while ensuring the safe and stable operation of the grid.

[0122] If the technical solution of this invention involves personal information, the product using this solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If sensitive personal information is involved, the user's separate consent has been obtained before processing, and the "express consent" requirement is met. For example, a clear sign is set at the collection device such as a camera to inform the user that they have entered the collection area, and the user's voluntary entry is considered as consent; or the processing device clearly indicates the processing rules and obtains authorization through pop-up windows or by asking the user to upload information themselves. The personal information processing rules include the processor, processing purpose, processing method, and types of personal information.

[0123] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power load forecasting method based on the fusion of fractional-order calculus control and graph neural network, characterized in that, The method includes: Based on historical load data and multi-source factor data, a preliminary trend feature vector of the time series is extracted through a long short-term memory network. The system dynamically decides whether to integrate sudden demand information based on whether the regional influence intensity exceeds a preset threshold. If it does, the sudden demand information is integrated to generate an enhanced trend vector. A convolutional neural network is used to extract spatial features from the enhanced trend vector. At the same time, based on the interaction between environmental influence and data quality, the dimension of the feature vector is dynamically adjusted to obtain optimized trend representation data. A weighted trend model is constructed by fusing the weight distributions of multiple factors through an attention mechanism. Based on the prediction results of the weighted trend model at different time scales, it is determined whether the prediction accuracy is lower than a preset threshold. If it is lower, the regional action parameters are adjusted back to determine the final prediction model. If the environmental impact fluctuates greatly, data quality filtering is automatically triggered to obtain stable load prediction results. Based on the stable load forecast results output by the final forecast model, a coordinated mapping between multiple regions is generated, and real-time verification is performed based on sudden changes in demand to output a comprehensive power load optimization scheme.

2. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network as described in claim 1, characterized in that, If the environmental impact fluctuates significantly, the data quality filtering will be automatically triggered to obtain a stable load forecast result, specifically including: The final prediction model is used to obtain a stable power supply simulation output. The model is then calibrated to address potential risks of resource waste and to determine whether environmental fluctuations are significant. If environmental fluctuations are significant, data quality filtering is automatically triggered to obtain a stable load prediction result.

3. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network as described in claim 1, characterized in that, The preliminary trend feature vector of the time series, extracted using a long short-term memory network based on historical load data and multi-source factor data, includes: Load-related records are retrieved from stored historical data, and the data is categorized and organized according to different time periods to obtain a structured load dataset. The load trends contained therein are analyzed, and statistical methods are used to identify the main change patterns and determine the key trend directions. Based on key trend directions and combined with data records from multiple sources, the dynamic correlation between the two is extracted to identify the main external variables affecting the trend. If the identified major external variables exceed the preset threshold range, the time series data is segmented to obtain the segmented sequence fragments, and a long short-term memory network is used for deep learning processing to generate the corresponding trend feature vectors. By combining the generated trend feature vector with the results of dynamic correlation, the load trend is further refined and decomposed to obtain multi-dimensional feature representation data. Then, the load changes of subsequent time series are predicted and simulated to obtain the predicted trend direction.

4. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network as described in claim 1, characterized in that, The step of dynamically determining whether to integrate sudden demand information based on whether the regional effect intensity exceeds a preset threshold, and if it does, then integrating the sudden demand information to generate an enhanced trend vector, includes: Obtain gridded load density data within the geographic range corresponding to the preliminary trend feature vector, and perform spatial aggregation calculation on the weighted average of the load values ​​of adjacent grids in the data to obtain load distribution data under regional action and its corresponding regional action intensity index. The regional action intensity index is compared with a preset threshold. If the regional action intensity index exceeds the preset threshold, the collection of sudden demand information within the synchronization time window is triggered. The sudden demand information comes from the external event report data stream. Principal component analysis is used to identify the demand pulse signal from the external event report data stream. The input of the principal component analysis is the numerical sequence of the event report data stream, and the output is the dimensionality-reduced demand pulse signal. The demand pulse signal is normalized, and the weight coefficient of the demand pulse signal in the fusion process is dynamically calculated based on the proportion of the regional effect intensity index exceeding the threshold and a fixed base. Through vector space mapping, the weighted demand pulse signal is converted into a demand influence vector with the same dimension as the preliminary trend feature vector, wherein the vector space mapping projects the signal to the corresponding dimension of the vector space. Within the vector space, the preliminary trend feature vector and the demand influence vector are superimposed. The superposition operation follows a linear weighting rule and is determined by adding elements one by one. The result is determined as the enhanced trend vector.

5. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network according to claim 1, characterized in that, The process involves using a convolutional neural network to extract spatial features from the enhanced trend vector, and dynamically adjusting the feature vector dimension based on the interaction between environmental influence and data quality to obtain optimized trend representation data. This includes: Obtain an enhanced trend vector, extract spatial sequence data from the trend vector, and perform convolution operation on the spatial sequence data using a convolutional neural network to obtain preliminary spatial feature representation data; For the preliminary spatial feature representation data, environmental impact factor data and data quality indicators are collected, and the two are multiplied to calculate the interaction coefficient. It is then determined whether the interaction coefficient exceeds a preset threshold to obtain the interaction judgment result. If the interaction judgment result indicates that adjustment is needed, then the preliminary spatial feature representation data is subjected to matrix multiplication dimension transformation according to the interaction coefficient to obtain the adjusted spatial feature representation data; The adjusted spatial feature representation data is used to perform vector weighted summation and fusion of the load distribution aggregate data and the sudden demand fusion information to calculate the grid load balance index and determine the dynamic demand calibration vector. The dynamic demand calibration vector is superimposed onto the adjusted spatial feature representation data to obtain the optimized trend representation data.

6. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network according to claim 1, characterized in that, The method of fusing the weight distributions of multiple factors through an attention mechanism to construct a weighted trend model includes: By obtaining historical samples of electricity load, sequence data is acquired to obtain initial trend representation data; An attention mechanism is used to fuse multiple factors, extract features from the initial trend representation data, and determine the weight distribution; Based on the weight distribution and combined with historical samples, the trend representation data is optimized to obtain the weighted trend; For the weighted trend, the difference between the peak and trough values ​​is calculated from the load sequence. If the difference exceeds a preset threshold, the parameters of the multi-source factors are adjusted, and the stability of the model is judged by comparing the trend fluctuations before and after the adjustment. The stability of the model is obtained, and load peak-valley regulation is integrated to construct a weighted trend model.

7. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network according to claim 1, characterized in that, The step involves iteratively updating the dynamic correlation based on the weighted trend model at different time scales, determining whether the prediction accuracy is lower than a preset threshold, and if so, backtracking to adjust the regional action parameters to determine the final prediction model, including: Obtain the prediction result sequence of the weighted trend model at three time scales: hourly, daily, and weekly, as well as the actual load observation sequence at the corresponding time scales; Based on the predicted result sequence and the observation sequence, calculate the prediction error sequence for each time scale; Using the Pearson correlation coefficient, the dynamic correlation between hourly scale error and daily scale error, and between daily scale error and weekly scale error, is analyzed to obtain a correlation strength matrix, which is used to describe the degree of mutual influence of prediction errors at different time scales. Based on the correlation strength matrix, the dynamic weight allocation mechanism in the weighted trend model is iteratively updated to adjust the model's dependence weight on historical data at different time scales; wherein, if the correlation strength between the hourly and daily scales is high, the weight of hourly historical data in daily scale prediction is increased. After iterative updates, the first updated model is obtained. Iterative optimization and convergence determination based on parameter backtracking mechanism are then performed to determine the final prediction model.

8. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network according to claim 7, characterized in that, After the iterative update, a first updated model is obtained. Iterative optimization and convergence determination based on a parameter backtracking mechanism are then performed to determine the final prediction model, including: Using the first update model, load forecasting is performed for a preset future time period to generate a second forecast sequence; The second predicted sequence is compared with the new actual observed sequence to calculate the overall prediction accuracy; Determine whether the overall prediction accuracy is lower than a preset accuracy threshold; If the accuracy is lower than the threshold, the parameter backtracking mechanism is triggered; The parameter backtracking mechanism targets specific regions with large prediction errors. Specifically, it extracts the predicted load curve for the region from the second prediction sequence and the actual load curve for the corresponding region from the actual observation sequence. It analyzes the differences between the two curves, decomposes the error components by point-by-point comparison, and identifies the systematic deviation mainly caused by insufficient characterization of regional load features. Based on the direction and magnitude of the systematic deviation, it adjusts the action parameter vector corresponding to the region in the first update model in reverse. Using the adjusted action parameter vector, it updates the first update model to obtain the second update model. Repeat the above prediction, accuracy judgment and parameter backtracking process until the overall prediction accuracy of three consecutive iterations is not lower than the preset accuracy threshold, or the adjustment range of the model parameters is less than a preset convergence threshold. Once the model is determined to be stable, the model parameters obtained from the last iteration are used as the parameters for the final prediction model.

9. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network according to claim 2, characterized in that, The process involves obtaining a stable power supply simulation output from the final prediction model, calibrating against potential risks of resource waste, determining the extent of environmental fluctuations, and, if so, incorporating data quality filtering to obtain a stable load prediction result. This includes: A stable analog output is obtained, and the simulation output is calibrated by comparing it with a preset risk threshold to address the potential risk of resource waste. For the calibrated analog output, it is determined whether the environmental influence fluctuation is large. If the environmental influence fluctuation is large, data quality filtering is incorporated, which is achieved by removing outliers, and the filtered output data is determined. From the filtered output data, the distribution characteristics of the power grid load balancing are obtained and determined by calculating the load variance, thus determining the balanced load data. For the balanced load data, environmental fluctuation calibration is incorporated by adjusting the fluctuation coefficient to obtain the fluctuation calibration result; From the fluctuation calibration results, the stable load prediction results are obtained.

10. The power load forecasting method based on the fusion of fractional-order calculus control and graph neural network according to claim 1, characterized in that, The process of generating a coordinated mapping between multiple regions based on the stable load forecast results output by the final forecast model, and verifying it in real time based on changes in sudden demand, and outputting a comprehensive power load optimization scheme includes: generating a coordinated mapping between multiple regions based on the stable load forecast results output by the final forecast model, performing preliminary integration on a preset regional energy sharing agreement, the agreement including inter-regional energy exchange rules, and obtaining a set of mapping relationships. The mapping relationship set is obtained, and real-time verification is performed for changes in sudden demand. If the verification result exceeds the preset threshold, the mapping parameters are adjusted by adding or subtracting parameters to obtain the verified mapping. Demand fluctuation response data is extracted from the verified mapping, and balance calculation is performed using a preset regional interaction relationship. The calculation is determined by a weighted summation method based on the interaction relationship to determine the power balance adjustment value. Based on the power balance adjustment value, the preset scheme adjustment logic and stability evaluation index are integrated. The integration is carried out by matching the index with logical rules to determine the direction of resource allocation optimization and obtain the optimized intermediate scheme. By integrating the load forecasting results with the real-time verification process through the optimized intermediate scheme, the comprehensive optimization scheme is finally calibrated. The calibration is achieved through result comparison and adjustment, and the comprehensive power load optimization scheme is output.

Images