A real-time power load prediction method and system based on a mixture model

Abstract

This application relates to a real-time power load forecasting method and system based on a hybrid model. The method includes: acquiring current cycle load data and combining it with historical load data to form a load sequence, and acquiring corresponding weather data; performing timestamp alignment, anomaly processing, and normalization on the load sequence and weather data to obtain a standardized input sequence; determining the order parameters of the linear forecasting model and the set of hyperparameters to be optimized, consisting of the network structure and training parameters of the nonlinear forecasting model, and optimizing them through a combined optimization algorithm to update the training configuration of the two models; outputting the first and second forecast sequences for the next cycle in the current cycle, respectively, and using the second forecast sequence as a trend term to compensate the first forecast sequence to obtain a fused forecast sequence; acquiring the fused forecast sequence for the current cycle from the previous cycle, constructing a residual sequence with the current cycle load data and determining the deviation term, calibrating the current cycle fused forecast sequence online, and outputting the calibrated load forecast result.

Description

Technical Field

[0001] This application relates to the field of power load forecasting technology, and in particular to a real-time power load forecasting method and system based on a hybrid model. Background Technology

[0002] Electricity load forecasting is a core component in ensuring the safe, stable operation and economically efficient dispatch of the power system. It is widely used in critical business scenarios such as grid dispatching decisions, electricity spot market clearing, demand response resource organization, and distribution network safety verification. With the continuous increase in the penetration rate of new energy sources such as wind and solar power, and the expanding scale of load-side adjustable resources such as distributed energy storage, electric vehicles, and flexible loads, coupled with drastic fluctuations in weather conditions and increased randomness in user electricity consumption behavior, electricity load data exhibits significantly enhanced randomness, nonlinearity, and non-stationarity. Load curves not only contain long-term trend components influenced by seasonal changes, weekly work patterns, and holiday effects, but also high-frequency fluctuation components caused by short-term weather changes, emergencies, and adjustments in industrial production rhythms. This complexity presents a significant challenge in obtaining stable, accurate, and continuously adaptive load forecasting results in real-time or near-real-time forecasting scenarios.

[0003] Existing load forecasting technologies are mainly divided into two categories: traditional statistical model methods and machine learning and deep learning methods. Traditional statistical models, such as autoregressive models (AR), moving average models (MA), and autoregressive integral moving average models (ARIMA), have advantages in handling linear relationships, periodic patterns, and long-term trends, including simple model structure, strong interpretability, and controllable computational resource consumption. However, their ability to characterize nonlinear fluctuations is significantly limited. When faced with rapid load changes driven by complex exogenous factors, such as extreme weather events or sudden load fluctuations, traditional models struggle to accurately capture the nonlinear dynamic characteristics of the data, leading to a significant increase in prediction errors. On the other hand, deep learning models, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), and sequence-to-sequence models (Seq2Seq), can learn complex nonlinear mapping relationships from multi-source data, exhibiting strong fitting capabilities for short-term fluctuations and sudden features. However, their model performance is highly dependent on network structure design and the selection of training parameters. Deep learning models are sensitive to hyperparameters and are prone to unstable generalization capabilities under different time scales, seasonal variations, or operational scenarios. Especially during real-time rolling prediction, due to the dynamic drift of data distribution, the limitations of model parameter update strategies, and the inadequacy of training set construction, prediction performance fluctuates greatly, making it difficult to maintain the reliability of long-term prediction. Summary of the Invention

[0004] To overcome, to some extent, the problems of existing load forecasting methods that cannot simultaneously take into account the linear trend and nonlinear fluctuation of load, and that rely on manual experience to adjust parameters, resulting in insufficient model generalization and robustness, this application provides a real-time power load forecasting method and system based on a hybrid model.

[0005] The proposed solution is as follows: According to a first aspect of the embodiments of this application, a real-time power load forecasting method based on a hybrid model is provided, comprising: The current period's power load data is obtained and combined with historical period's power load data to form a power load sequence. At the same time, the weather data corresponding to the power load sequence is obtained. The power load sequence and the weather data are subjected to timestamp alignment, anomaly handling and normalization processing to obtain a standardized input sequence. Determine the order parameters of the linear prediction model and the set of hyperparameters to be optimized, consisting of the network structure parameters and training parameters of the nonlinear prediction model; The set of hyperparameters to be optimized is optimized by a combinatorial optimization algorithm, and the linear prediction model and the nonlinear prediction model are trained and updated using the optimized set of hyperparameters. The standardized input sequence is input into the nonlinear prediction model and the linear prediction model respectively. The nonlinear prediction model outputs a first prediction sequence, and the linear prediction model outputs a second prediction sequence. The prediction sequence is the power load prediction sequence for the next cycle in the current cycle. The first prediction sequence is compensated using the second prediction sequence as a trend term to obtain the fused prediction sequence; Obtain the fused prediction sequence for the current period from the output of the previous period. Calculate the residuals based on the fused prediction sequence of the previous period and the power load data of the current period, and construct a residual sequence. Determine the deviation term based on the residual sequence. Use the deviation term to perform online calibration on the fused prediction sequence of the current period, and output the calibrated load prediction result.

[0006] Preferably, before inputting the standardized input sequence into the nonlinear prediction model and the linear prediction model respectively, the method further includes: Based on the standardized input sequence, a training set and a validation set are constructed to determine the set of hyperparameters to be optimized; the set of hyperparameters to be optimized includes at least: the order parameters of the linear prediction model, and the network structure parameters and training parameters of the nonlinear prediction model; The set of hyperparameters to be optimized is encoded to construct an initial population for a genetic algorithm, and the prediction error on the validation set is used as the fitness function. The initial population is iteratively updated through selection, crossover, and mutation to obtain a set of candidate optimal hyperparameters. Using the candidate set of preferred hyperparameters as the initial positions of particles in the particle swarm optimization algorithm, the particle velocity and position are updated based on the individual optimal solution and the global optimal solution, and the target hyperparameter combination is obtained iteratively. Based on the target hyperparameter combination, the nonlinear prediction model and the linear prediction model are updated with parameters and training configuration, and the updated models are used to output the first prediction sequence and the second prediction sequence, respectively.

[0007] Preferably, the method further includes: During real-time rolling prediction, the training parameters are adaptively adjusted based on the changes in the prediction residuals within the most recent preset window. The training parameters include at least the learning rate. Select a hidden layer size that matches the current residual level from the preset network size candidate set in order to maintain the adaptability of the nonlinear prediction model to load fluctuations at different time scales.

[0008] Preferably, before performing normalization processing on the power load sequence and the weather data, the method further includes: A time feature sequence is generated based on the timestamp of the power load sequence; the time feature sequence includes time feature, weekday feature and holiday identifier; The time feature sequence and the weather data are used together as exogenous features to form a multivariate input sequence with the power load sequence.

[0009] Preferably, compensating the first prediction sequence with the second prediction sequence as a trend term includes: The first prediction sequence and the second prediction sequence are timestamp aligned according to a preset prediction step size to obtain an aligned prediction sequence. Scale matching is performed on the second prediction sequence in the aligned prediction sequence so that the amplitude scale of the second prediction sequence is consistent with that of the first prediction sequence; The trend compensation parameters are determined based on the residual sequence within the most recent preset window; the most recent preset window is the time window between the previous period and the current period; the trend compensation parameters include at least a trend compensation coefficient and a bias term. Based on the trend compensation parameters, a trend compensation term is generated for the second prediction sequence, and the trend compensation term is superimposed on the first prediction sequence to obtain the fused prediction sequence.

[0010] Preferably, the trend compensation coefficient is a dynamic coefficient; the dynamic coefficient is updated according to the weather characteristics, time period characteristics or residual level of the current period to adjust the influence of the trend compensation term on the fused prediction sequence.

[0011] Preferably, the dynamic coefficients are updated based on the weather characteristics, time period characteristics, or residual level of the current period, including: Obtain the weather feature vector, time period feature vector, and residual sequence within the most recent preset window for the current period, and calculate the residual level index based on the residual sequence within the most recent preset window; the residual level index includes at least one or more of the residual mean, residual root mean square, or residual absolute mean. The weather feature vector, the time period feature vector, and the residual level index are input into the dynamic coefficient update function to obtain candidate values ​​for the dynamic coefficient. Boundary constraint processing is performed on the candidate values ​​to ensure that the dynamic coefficients are within a preset range; The dynamic coefficients are updated based on a smooth update strategy, which includes: weighting and fusing the dynamic coefficients of the previous period with the candidate values ​​at a preset update step size to obtain the dynamic coefficients of the current period. When the residual level index exceeds a preset threshold, the dynamic coefficient is reduced to decrease the impact of the trend compensation term; when the residual level index is below the preset threshold, the dynamic coefficient is increased to enhance the impact of the trend compensation term.

[0012] Preferably, the residual is calculated based on the fused prediction sequence of the previous period and the power load data of the current period, and a residual sequence is constructed. The deviation term is determined based on the residual sequence, including: The fused prediction sequence for the current period output from the previous period is timestamped with the power load data of the current period to obtain an aligned sample. Calculate the residuals based on the aligned samples and construct the residual sequence; The bias term is obtained by weighting the residual sequence according to the time decay weight. The deviation term is superimposed on the fusion prediction sequence of the current period to complete the online calibration.

[0013] Preferably, the method further includes: During real-time rolling forecasting, the drift index is calculated based on the residual sequence within the most recent preset window. When the drift metric meets the preset triggering condition, the rolling window data of the training set and validation set is updated, and the combined optimization algorithm is triggered to re-optimize the set of hyperparameters to be optimized, so as to update the training configuration of the linear prediction model and the nonlinear prediction model.

[0014] According to a second aspect of the embodiments of this application, a real-time power load forecasting system based on a hybrid model is provided, comprising: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute a real-time power load forecasting method based on a hybrid model as described in any of the above.

[0015] The technical solution provided in this application may include the following beneficial effects: This technical solution aligns load data and weather data with timestamps, handles anomalies, and normalizes them to ensure temporal consistency and data quality across multiple input sources. It further determines the hyperparameter set consisting of the linear model order, the nonlinear model network structure, and training parameters, and optimizes it using a combinatorial optimization algorithm. This allows the training configurations of the two models to achieve better matching in different scenarios, reduces reliance on manual parameter tuning, and improves prediction stability. Based on this, it outputs linear and nonlinear prediction sequences respectively, using the linear prediction sequence as a trend term to compensate for the nonlinear prediction sequence, achieving collaborative modeling of linear trends and nonlinear fluctuations in load. Simultaneously, it constructs a residual sequence using the fusion prediction of the previous period and the actual load of the current period, extracts the deviation term, and performs online calibration of the fusion prediction for the current period, thereby suppressing the accumulation of systematic deviations and enhancing the accuracy and robustness of real-time rolling predictions.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 This is a flowchart illustrating a real-time power load forecasting method based on a hybrid model, provided in one embodiment of this application. Figure 2 This is a schematic diagram of the structure of a real-time power load forecasting system based on a hybrid model, provided in one embodiment of this application.

[0019] Reference numerals: Processor-21; Memory-22. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0021] Traditional power load forecasting technologies, whether statistical or deep learning models, struggle to simultaneously achieve stable, accurate, and sustainably adaptive load forecasts in real-time or near-real-time scenarios. Statistical models are inadequate at characterizing nonlinear fluctuations, while deep learning models are sensitive to hyperparameters and prone to generalization instability and performance fluctuations in real-time rolling forecasting scenarios.

[0022] To address this, this application provides a real-time power load forecasting method based on a hybrid model, referring to... Figure 1 ,include: S1. Obtain the current period's power load data and combine it with the historical period's power load data to form a power load sequence. At the same time, obtain the weather data corresponding to the power load sequence. Perform timestamp alignment, anomaly handling, and normalization on the power load sequence and weather data to obtain a standardized input sequence. S2. Determine the order parameters of the linear prediction model and the set of hyperparameters to be optimized, which consists of the network structure parameters and training parameters of the nonlinear prediction model. S3. Optimize the set of hyperparameters to be optimized using a combinatorial optimization algorithm, and then use the optimized set of hyperparameters to train and update the configuration of the linear and nonlinear prediction models. S4. Input the standardized input sequence into the nonlinear prediction model and the linear prediction model respectively. Output the first prediction sequence through the nonlinear prediction model and the second prediction sequence through the linear prediction model. The prediction sequence is the power load prediction sequence for the next cycle in the current cycle. S5. The first prediction sequence is compensated using the second prediction sequence as the trend term to obtain the fused prediction sequence; S6. Obtain the fused prediction sequence for the current period from the previous period's output. Calculate the residuals and construct a residual sequence based on the fused prediction sequence for the previous period and the power load data for the current period. Determine the deviation term based on the residual sequence. Use the deviation term to perform online calibration on the fused prediction sequence for the current period. Output the calibrated load prediction result.

[0023] For ease of understanding, the following explains some key terms in this embodiment: An electricity load sequence refers to a set of electricity consumption data obtained through monitoring or recording within a specific time period. This sequence is usually indexed by time and reflects the pattern of electricity load changes over time.

[0024] Weather data refers to meteorological information related to power load forecasting, such as temperature, humidity, wind speed, and rainfall. This data is used to capture the impact of weather conditions on power load.

[0025] A standardized input sequence refers to multivariate time series data that has undergone preprocessing steps such as timestamp alignment, anomaly handling, and normalization. This sequence serves as input to the predictive model, ensuring data quality and the stability of model training.

[0026] Linear forecasting models are mathematical models that can capture the linear trends and periodic characteristics of electricity load data, such as autoregressive (AR) models, moving average (MA) models, or autoregressive moving average (ARMA) models. These models predict future load by linearly combining historical data.

[0027] Nonlinear prediction models refer to machine learning or deep learning models that can capture the complex nonlinear relationships and high-frequency fluctuations in power load data, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or gated recurrent units (GRUs). These models learn the relationship between inputs and outputs through nonlinear mapping.

[0028] The set of hyperparameters to be optimized refers to the set of parameters that need to be pre-defined before model training, and their values ​​directly affect the model's performance. For linear prediction models, this set may include order parameters; for nonlinear prediction models, this set may include network structure parameters and training parameters such as the number of network layers, number of nodes, learning rate, and batch size.

[0029] Combinatorial optimization algorithms are optimization methods used to find optimal solutions under given constraints, such as genetic algorithms (GA), particle swarm optimization (PSO), or hybrid forms thereof. These algorithms are used to efficiently search the hyperparameter space to find the hyperparameter combination that optimizes model performance.

[0030] The first prediction sequence refers to the prediction results of future power load output by a nonlinear prediction model based on a standardized input sequence. This sequence mainly reflects the nonlinear fluctuation characteristics of the load.

[0031] The second forecast sequence refers to the forecast results of future electricity load output by the linear forecasting model based on the standardized input sequence. This sequence mainly reflects the linear trend characteristics of the load.

[0032] A fusion prediction sequence refers to a comprehensive prediction result obtained by compensating the first prediction sequence with a second prediction sequence as a trend term. This sequence combines the trend-capturing ability of a linear model with the fluctuation-fitting ability of a nonlinear model.

[0033] The residual sequence is the sequence of differences between actual power load data and model predictions. This sequence reflects the model's prediction error and is used to evaluate model performance and for subsequent calibration.

[0034] The bias term refers to the correction amount calculated from the residual sequence and used to correct the current period prediction result. This term aims to correct systematic biases in the model prediction.

[0035] Online calibration refers to the process of dynamically adjusting and correcting the forecast results for the current period using the latest actual load data and forecast residuals during real-time rolling forecasting. This process enhances the real-time adaptability and accuracy of the forecast results.

[0036] This embodiment provides a real-time power load forecasting method based on a hybrid model. Specifically, the method first acquires the power load data for the current period and combines it with power load data from historical periods to form a power load sequence, while simultaneously acquiring the weather data corresponding to the power load sequence. Subsequently, timestamp alignment, anomaly handling, and normalization are performed on the power load sequence and the weather data to obtain a standardized input sequence. For example, power load data can be obtained from a power dispatching system or smart meters, and historical data can be stored in a database. Weather data can be obtained from a meteorological bureau interface, a third-party meteorological service platform, or a local weather station. Timestamp alignment can be achieved through a time-index-based merging operation. Outliers can be identified and removed using statistical methods, and missing data can be filled using interpolation or mean filling. Normalization can employ min-max normalization or Z-score standardization to eliminate dimensional differences.

[0037] Further, the order parameter of the linear prediction model and the network structure parameters and training parameters of the nonlinear prediction model are determined to form a set of hyperparameters to be optimized. For example, the order parameter of the linear prediction model can be set empirically to a fixed length of historical data. The network structure parameters of the nonlinear prediction model can be initially set to common network layer and node numbers, and the training parameters can be set to default learning rate and batch size. Subsequently, this set of hyperparameters to be optimized is optimized using a combinatorial optimization algorithm. This combinatorial optimization algorithm can employ a heuristic search algorithm, such as random search or grid search, to iteratively try different combinations of hyperparameters within a predefined hyperparameter space and evaluate the model performance under different hyperparameter combinations. Through multiple iterations, the hyperparameter combination with better performance is selected. Thus, the training configuration of the linear and nonlinear prediction models is updated using the optimized hyperparameter set. For example, the order of the linear model is set based on the optimization results, or the number of layers, nodes, and learning rate of the nonlinear model are adjusted.

[0038] Based on this, the standardized input sequence is input into both the nonlinear and linear prediction models. Specifically, the preprocessed standardized input sequence, such as multidimensional data including historical load and weather characteristics, is simultaneously provided to both the trained nonlinear and linear prediction models. The nonlinear prediction model outputs a first prediction sequence, and the linear prediction model outputs a second prediction sequence. The prediction sequences are power load prediction sequences for the next cycle based on the current cycle. For example, after receiving the input, the nonlinear prediction model generates the prediction result for future load—the first prediction sequence—through its internal nonlinear mapping relationship. Similarly, the linear prediction model, after receiving the input, generates the prediction result for future load—the second prediction sequence—through its linear relationship modeling capability. The prediction period can be the power load prediction for the next hour or day from the current moment.

[0039] In a preferred implementation, the first predicted sequence is compensated using the second predicted sequence as a trend term, thereby obtaining a fused predicted sequence. Specifically, the second predicted sequence output by the linear prediction model is directly used as the trend term. The compensation operation can be a simple weighted sum of the first and second predicted sequences, or the second predicted sequence can be directly superimposed on the first predicted sequence to correct any trend deviations that may exist in the first predicted sequence. Through the above compensation operation, a final prediction result combining linear trends and nonlinear fluctuations is obtained, i.e., the fused predicted sequence.

[0040] Furthermore, the fused prediction sequence for the current period from the previous period's output is obtained. Residuals are calculated based on the fused prediction sequence from the previous period and the current period's power load data, and a residual sequence is constructed. For example, the predicted value for the current time from the previous period's model is obtained and compared with the actual power load data for the current time, and the difference between the two is calculated to form a residual. The residual values ​​from multiple consecutive time steps are arranged in chronological order to form a residual sequence. A deviation term is determined based on this residual sequence. For example, a simple average of the residual sequence can be performed, or the most recent residual value can be taken as the deviation term. Finally, this deviation term is used to perform online calibration on the fused prediction sequence for the current period, and the calibrated load prediction result is output. For example, the determined deviation term can be directly added to the fused prediction sequence output by the current period's model to correct the prediction result in real time.

[0041] This technical solution aligns load data and weather data with timestamps, handles anomalies, and normalizes them to ensure temporal consistency and data quality across multiple input sources. It further determines the hyperparameter set consisting of the linear model order, the nonlinear model network structure, and training parameters, and optimizes it using a combinatorial optimization algorithm. This allows the training configurations of the two models to achieve better matching in different scenarios, reduces reliance on manual parameter tuning, and improves prediction stability. Based on this, it outputs linear and nonlinear prediction sequences respectively, using the linear prediction sequence as a trend term to compensate for the nonlinear prediction sequence, achieving collaborative modeling of linear trends and nonlinear fluctuations in load. Simultaneously, it constructs a residual sequence using the fusion prediction of the previous period and the actual load of the current period, extracts the deviation term, and performs online calibration of the fusion prediction for the current period, thereby suppressing the accumulation of systematic deviations and enhancing the accuracy and robustness of real-time rolling predictions.

[0042] Example 2 It should be noted that before inputting the standardized input sequence into the nonlinear prediction model and the linear prediction model respectively, the following steps are also included: Training and validation sets are constructed based on standardized input sequences to determine the set of hyperparameters to be optimized. The set of hyperparameters to be optimized includes at least the order parameters of the linear prediction model and the network structure parameters and training parameters of the nonlinear prediction model. The set of hyperparameters to be optimized is encoded, an initial population for a genetic algorithm is constructed, and the prediction error on the validation set is used as the fitness function. The initial population is iteratively updated through selection, crossover, and mutation to obtain a set of candidate optimal hyperparameters. The candidate set of optimal hyperparameters is used as the initial position of particles in the particle swarm optimization algorithm. The particle velocity and position are updated based on the individual optimal solution and the global optimal solution, and the target hyperparameter combination is obtained iteratively. Based on the target hyperparameter combination, the parameters and training configuration of the nonlinear prediction model and the linear prediction model are updated, and the updated models are used to output the first prediction sequence and the second prediction sequence, respectively.

[0043] To efficiently find the optimal hyperparameter combination, this application employs a combinatorial optimization strategy. Specifically, the set of hyperparameters to be optimized is encoded to construct an initial population for a genetic algorithm, and the prediction error on the validation set is used as the fitness function. The initial population is iteratively updated through selection, crossover, and mutation operations to obtain a set of candidate optimal hyperparameters. In this process, encoding converts the hyperparameter set into a gene sequence or chromosome form operable by the genetic algorithm, and the initial population is a randomly generated set of hyperparameter combinations. The fitness function is used to evaluate the quality of each hyperparameter combination, and is usually represented by the reciprocal or negative value of the model's prediction error on the validation set (e.g., mean squared error, mean absolute error, etc.). The smaller the error, the higher the fitness. Selection, crossover, and mutation are the core operations of the genetic algorithm. By simulating natural selection and genetic mechanisms, individuals in the population continuously evolve, gradually approaching the optimal solution, thereby obtaining a set of candidate optimal hyperparameters with better performance.

[0044] Based on this, the candidate set of optimal hyperparameters is used as the initial particle positions for the Particle Swarm Optimization (PSO) algorithm. Particle velocities and positions are updated based on individual optimal solutions and the global optimal solution, iteratively obtaining the target hyperparameter combination. Particle Swarm Optimization (PSO) is a swarm intelligence-based optimization algorithm that simulates the foraging behavior of bird flocks. Using the candidate set of optimal hyperparameters obtained from a genetic algorithm as the initial particle positions for PSO provides a good starting point and avoids getting trapped in local optima. Each particle represents a hyperparameter combination. By tracking the individual historical best solution (pbest) and the swarm's global best solution (gbest), the particle adjusts its velocity and position, moving within the search space to find a better hyperparameter combination. This combinatorial optimization strategy effectively improves search efficiency and global optimization capability.

[0045] Finally, based on the target hyperparameter combination, the nonlinear prediction model and the linear prediction model are updated in terms of parameters and training configuration. The updated models are then used to output the first prediction sequence and the second prediction sequence, respectively. The target hyperparameter combination is the optimal hyperparameter configuration obtained after optimization through a combinatorial optimization algorithm. Based on this combination, the linear and nonlinear prediction models are updated in terms of parameters and training configuration. This means that the models will be retrained or configured using these optimized hyperparameters to ensure optimal performance in the prediction task. Subsequently, using these optimized models, the first and second prediction sequences are generated, respectively, providing more accurate input for subsequent fusion and calibration steps.

[0046] Through the above technical solution, this application can efficiently and accurately determine the key hyperparameters of linear and nonlinear prediction models. Specifically, the global search capability of the genetic algorithm can effectively explore the hyperparameter space, avoid getting trapped in local optima, and generate a set of potential candidate hyperparameters. Based on this, the particle swarm optimization algorithm uses these candidate sets as initial positions to further refine the search, accelerate convergence, and improve optimization accuracy. This two-stage combined optimization method overcomes the limitations that may exist in a single optimization algorithm, ensuring that the determined target hyperparameter combination enables the hybrid prediction model to fully learn data features on the training set and exhibit good generalization ability and prediction accuracy on the validation set. Finally, the model is configured and trained based on the optimized hyperparameters, resulting in higher initial quality of the first and second prediction sequences, providing a more solid foundation for subsequent fusion compensation and online calibration, thereby significantly improving the accuracy and stability of overall power load prediction.

[0047] Furthermore, the methods also include: During real-time rolling prediction, the training parameters are adaptively adjusted based on the changes in the prediction residuals within the most recent preset window. The training parameters include at least the learning rate. Select a hidden layer size that matches the current residual level from the preset candidate set of network sizes to maintain the adaptability of the nonlinear prediction model to load fluctuations at different time scales.

[0048] Specifically, in real-time rolling prediction, the model needs to continuously predict and adjust based on the latest data. Training parameters, such as the learning rate, are key factors affecting the model's convergence speed and final performance. When the fluctuation pattern of power load or the external environment changes, a fixed learning rate may lead to low model training efficiency, or even overfitting or underfitting. Therefore, adaptively adjusting the training parameters based on the changes in the prediction residuals within the most recent preset window means that the system continuously monitors the model's prediction error (i.e., prediction residuals) over a recent period and dynamically adjusts the training parameters according to the trends of these residuals. For example, when the prediction residuals continue to increase or fluctuate drastically, it may mean that the current learning rate of the model is too high, causing the model to oscillate or fail to converge to the optimal solution. In this case, the system can automatically reduce the learning rate to stabilize the training process; conversely, when the residuals are small and stable, the learning rate can be appropriately increased to accelerate the model's learning of new patterns. This adaptive adjustment mechanism ensures that the model can continuously and effectively learn and update in a dynamic environment.

[0049] By employing the aforementioned technical solution, during real-time rolling prediction, training parameters (such as the learning rate) are adaptively adjusted based on changes in the prediction residuals. Furthermore, the size of the hidden layer is dynamically selected according to the current residual level. This application enables the nonlinear prediction model to maintain continuous adaptability and robustness when facing dynamic changes in power load data and complex fluctuations at different time scales. This effectively solves the problem of performance degradation that may occur with preset fixed parameters and network structures over long-term operation, ensuring that the model can capture the latest load characteristics in real time, thereby significantly improving the accuracy and stability of real-time power load prediction.

[0050] Example 3 In some of the above implementations, the power load forecasting method mainly relies on power load sequences and weather data for prediction, and performs timestamp alignment, anomaly handling, and normalization on these data to obtain a standardized input sequence. However, power load data has significant time periodicity, seasonality, and is affected by specific dates (such as holidays). If only the raw load data and weather data are used as input, the model may struggle to fully capture these complex temporal patterns, thus affecting the accuracy of the forecast and the model's generalization ability.

[0051] In this regard, this application further proposes that, before performing normalization processing on the power load sequence and weather data, the following steps are also included: A time feature sequence is generated based on the timestamps of the power load sequence; the time feature sequence includes time feature, weekday feature and holiday identifier; The time feature series and weather data are used together as exogenous features to form a multivariate input series with the power load series.

[0052] Specifically, a time feature sequence is generated based on the timestamps of the power load sequence, aiming to extract periodic and categorical features valuable for power load prediction from the raw time information. These features help the model understand the inherent patterns of load changes over time, such as daily, weekly, and annual cycles. This can be achieved by parsing the timestamps of each data point in the power load sequence and extracting different time-dimensional information. For example, for a timestamp "2023-10-26 14:30", "hour" (14), "day of the week" (Thursday), "month" (10), and "year" (2023) can be extracted. This extracted information can be further encoded into numerical or categorical features.

[0053] The time feature sequence includes time-of-day features, weekday features, and holiday identifiers. Time-of-day features refer to specific points in time within a day, such as hours or minutes. Electricity load typically exhibits significant periodic variations throughout the day, such as peak consumption during the day and off-peak consumption at night. Introducing time-of-day features helps the model capture this intraday periodicity. This can be achieved by directly using hours (0-23) or minutes (0-59) as numerical features, or by using one-hot encoding to convert them into categorical features. Weekday features refer to specific dates within a week, such as Monday to Sunday. Electricity load typically shows significant differences between weekdays and weekends, for example, higher load on weekdays and lower load on weekends. Introducing weekday features helps the model distinguish these intraweekly periodicities. This can be achieved by directly using the day of the week (1-7) as numerical features, or by using one-hot encoding to convert them into a 7-dimensional binary vector. Holiday identifiers indicate whether a date is a holiday. During holidays, social activity patterns change, often resulting in significant differences in electricity load compared to ordinary weekdays or weekends. Introducing holiday identifiers helps the model identify and adapt to this special date effect. This can be achieved by querying a preset holiday calendar, assigning a binary identifier to each date (e.g., 1 for a holiday, 0 for a non-holiday), or by further dividing holidays into different types and performing one-hot encoding.

[0054] The time feature sequence and the weather data are used together as exogenous features to form a multivariate input sequence with the power load sequence. Exogenous features refer to external variables that are independent of the predicted variable (power load) but have an influence on it. In the data preprocessing stage, the power load sequence, the generated time feature sequence (including time features, weekday features, and holiday identifiers), and the weather data (such as temperature, humidity, wind speed, etc.) are aligned in the time dimension. Then, these aligned data are merged into a multidimensional input matrix or tensor, which serves as the input for subsequent linear and nonlinear prediction models. For example, for each time step, the input can be a vector containing the power load value for that time step, the corresponding time, weekday, holiday identifier, and various weather indicators.

[0055] By explicitly generating and incorporating time-based features, weekday features, and holiday identifiers, the model can directly learn the periodic variation patterns of electricity load at different time scales. Examples include intraday peak and off-peak times, differences between weekdays and weekends, and the specific impact of holidays on load. These time features, along with weather data, serve as exogenous features, forming a multivariate input sequence together with the electricity load series, greatly enriching the model's input information. This allows linear and nonlinear prediction models to consider not only historical load and weather factors but also to fully utilize temporal context information, thereby more accurately capturing complex load fluctuation patterns and significantly improving prediction accuracy and model generalization ability.

[0056] Example 4 It should be noted that the compensation for the first predicted sequence using the second predicted sequence as a trend term includes: The first prediction sequence and the second prediction sequence are aligned with the timestamps according to a preset prediction step size to obtain the aligned prediction sequence. Scale matching is performed on the second prediction sequence in the aligned prediction sequence to make the amplitude scale of the second prediction sequence consistent with that of the first prediction sequence; The trend compensation parameters are determined based on the residual sequence within the most recent preset window; the most recent preset window is the time window between the previous period and the current period; the trend compensation parameters include at least the trend compensation coefficient and the bias term. A trend compensation term is generated for the second prediction sequence based on the trend compensation parameter, and the trend compensation term is superimposed on the first prediction sequence to obtain the fused prediction sequence.

[0057] Specifically, aligning the first and second prediction sequences with timestamps according to a preset prediction step size aims to ensure strict consistency in the time dimension between the prediction results from the nonlinear and linear prediction models. This means that for each prediction point in time, the data in both sequences should correspond to the same future time. For example, if the preset prediction step size is 15 minutes, the first data point in both sequences should represent the load prediction for the next 15 minutes, the second data point for the 15 minutes after that, and so on. This alignment operation can be achieved by checking and adjusting the start and end times of the sequences or by interpolation, truncation, etc., to eliminate time offsets that may be caused by differences in model internal processing or data generation mechanisms.

[0058] Subsequently, scaling is performed on the second prediction sequence in the aligned prediction sequence to ensure that its amplitude scale is consistent with that of the first prediction sequence. Since nonlinear and linear prediction models may have different modeling principles, output characteristics, or sensitivities to input data, their directly output predictions may differ in numerical range or fluctuation amplitude. Scaling adjustment adjusts the amplitude of the second prediction sequence by applying appropriate transformations (e.g., linear scaling, standardization, or mapping functions learned from historical data) to make it numerically comparable to the first prediction sequence. This avoids over-amplification or under-amplification of the influence of one model due to scaling inconsistencies during subsequent compensation.

[0059] Based on this, trend compensation parameters are determined using the residual sequence within the most recent preset window. The most recent preset window typically covers the time window between the previous and current periods, such as the residual between the actual load data and the corresponding predicted values ​​for the most recent 24 hours or week. The residual sequence reflects the model's error in recent predictions. Trend compensation parameters include at least a trend compensation coefficient and a bias term, which are used to dynamically adjust the strength and direction of the compensation from the second prediction sequence to the first prediction sequence. For example, these parameters can be calculated in real time using statistical regression methods (such as least squares) or adaptive algorithms by analyzing the mean, variance, or trend of the residual sequence. The trend compensation coefficient can be a multiplicative factor used to adjust the relative contribution of the second prediction sequence; the bias term is an additive factor used to correct systematic prediction biases.

[0060] Finally, a trend compensation term is generated for the second prediction sequence based on the determined trend compensation parameters, and this trend compensation term is superimposed on the first prediction sequence to obtain the fused prediction sequence. The generation of the trend compensation term typically involves multiplying the second prediction sequence by the trend compensation coefficient and adding a bias term. By superimposing this finely tuned trend compensation term onto the first prediction sequence, this application effectively combines the trend-capturing capability of a linear model with the complex pattern recognition capability of a nonlinear model.

[0061] Preferably, the trend compensation coefficient is a dynamic coefficient; the dynamic coefficient is updated according to the weather characteristics, time period characteristics or residual level of the current period to adjust the influence of the trend compensation term on the fused prediction sequence.

[0062] Specifically, the trend compensation coefficient is set as a dynamic coefficient, meaning it is not fixed but can be adjusted in real time according to specific conditions or inputs. This dynamism gives the system the ability to flexibly adjust the intensity of trend compensation under different operating environments, thus better adapting to the complex changes in power load. The dynamic coefficient update mechanism is based on multi-dimensional information. Weather characteristics include, but are not limited to, temperature, humidity, wind speed, and precipitation, which have a significant impact on power load. By considering weather characteristics, the system can capture load trend fluctuations caused by weather changes and adjust the trend compensation accordingly. Time-period characteristics cover time dimensions such as hour, day, week, month, season, and holidays. Power load exhibits periodic patterns in different time periods, and the dynamic coefficient can be adjusted according to these time-period characteristics to reflect the expected contribution of the trend term in different time periods. The residual level refers to the magnitude of the error generated by the model in past predictions. When the residual level is high, it may mean that the model has a bias in capturing the current load trend. In this case, adjusting the dynamic coefficient can help correct this bias, for example, by strengthening or weakening the impact of the trend compensation term. The introduction of these characteristics makes the adjustment of the dynamic coefficient well-founded and can more accurately reflect the current prediction environment and model performance. The updating of the dynamic coefficients directly affects the trend compensation term, altering its value and thus changing the weight or magnitude of the trend compensation term when superimposed on the first prediction sequence. For example, when the dynamic coefficients increase, the influence of the trend compensation term on the fused prediction sequence strengthens; conversely, when the dynamic coefficients decrease, its influence weakens. This adjustment mechanism allows the system to finely control the role of the linear prediction model in the final fusion result, ensuring it achieves optimal compensation performance under different scenarios.

[0063] It should be noted that the dynamic coefficients are updated based on the weather characteristics, time period characteristics, or residual level of the current cycle, including: Obtain the weather feature vector, time period feature vector, and residual sequence within the most recent preset window for the current period, and calculate the residual level index based on the residual sequence within the most recent preset window; the residual level index includes at least one or more of the residual mean, residual root mean square, or residual absolute mean. The weather feature vector, time period feature vector, and residual level index are input into the dynamic coefficient update function to obtain candidate values ​​for the dynamic coefficient. Boundary constraints are applied to the candidate values ​​to ensure that the dynamic coefficients are within a preset range. The dynamic coefficients are updated based on a smooth update strategy, which includes: weighting and fusing the dynamic coefficients of the previous period and the candidate values ​​with a preset update step size to obtain the dynamic coefficients of the current period. When the residual level index exceeds the preset threshold, the dynamic coefficient is reduced to decrease the impact of the trend compensation term; when the residual level index is below the preset threshold, the dynamic coefficient is increased to enhance the impact of the trend compensation term.

[0064] Through the above technical solution, this application enables refined and adaptive updating of dynamic coefficients. By comprehensively considering external environmental factors such as weather and time period, as well as the model's own prediction residual level, and combining boundary constraints and smoothing update strategies, it ensures that the updating of dynamic coefficients can both quickly respond to environmental changes and maintain the stability and rationality of the update process. This intelligent dynamic coefficient adjustment mechanism allows the trend compensation term to act more accurately on the first prediction sequence, effectively solving the problem of untimely or unstable dynamic coefficient updates in traditional methods. This significantly improves the accuracy and robustness of the hybrid model in real-time power load forecasting, especially under conditions of large load fluctuations or complex and volatile external environments, providing more reliable prediction results.

[0065] Example 5 It should be noted that the residuals are calculated and a residual sequence is constructed based on the fused forecast sequence of the previous period and the power load data of the current period. The deviation term is determined based on the residual sequence, including: The fused prediction sequence for the current period output from the previous period is timestamped with the power load data of the current period to obtain an aligned sample. Calculate residuals based on aligned samples and construct residual sequences; The bias term is obtained by weighting the residual sequence according to the time decay weight. The bias term is superimposed onto the fusion prediction sequence of the current period to complete the online calibration.

[0066] Through the above technical solution, this application effectively addresses the problem of difficulty in effectively capturing and correcting persistent systematic biases or drifts that may exist in the model during real-time forecasting. Specifically, by aligning the timestamp of the fused forecast sequence for the current period output from the previous period with the power load data of the current period, the accuracy of residual calculation is ensured. Based on this, a weighted average of the residual sequence is applied according to time decay weights, enabling the determined bias term to dynamically and sensitively reflect the recent forecast error trend of the model, assigning greater correction weight to recent errors, thereby effectively capturing potential drifts or systematic biases in the model. Finally, this dynamic bias term is superimposed on the fused forecast sequence for the current period, achieving online calibration of the forecast results. This allows the calibrated load forecast results to adapt promptly to actual changes in power load, significantly improving the accuracy and stability of real-time power load forecasting, especially providing more reliable forecast support when load patterns undergo minor or persistent changes.

[0067] Furthermore, the methods also include: During real-time rolling forecasting, the drift index is calculated based on the residual sequence within the most recent preset window. When the drift metric meets the preset trigger conditions, the rolling window data of the training and validation sets is updated, and the combinatorial optimization algorithm is triggered to re-optimize the set of hyperparameters to be optimized, so as to update the training configuration of the linear prediction model and the nonlinear prediction model.

[0068] This drift index can be characterized by various statistical measures. For example, it can be the changing trend of the mean, variance, root mean square error (RMSE), and mean absolute error (MAE) of the residual series within a preset time window, or it can be detected by statistical hypothesis testing (such as the Page-Hinkley test, CUSUM control charts, etc.) to detect significant changes in the residual distribution. By continuously monitoring the residual series and calculating the index, the accumulation or change of model prediction error can be quantified, thus providing a basis for subsequent adaptive adjustments.

[0069] When the drift metric meets a preset trigger condition, the system will automatically update the rolling window data of the training and validation sets. When the calculated drift metric reaches or exceeds the preset trigger condition, it indicates that the model may have experienced performance degradation or a significant change in data distribution, requiring model retraining. In this case, to enable the model to adapt to the new data distribution, the datasets used for model training and validation need to be updated. Specifically, the training and validation sets are updated using a rolling window approach, removing the oldest data and incorporating the latest electricity load and weather data. This rolling window update mechanism ensures that the training and validation sets always contain the latest and most representative data, enabling the model to learn the latest data patterns and trends, avoiding prediction bias caused by using outdated data. The preset trigger condition can be a fixed threshold, such as when the drift metric (e.g., RMSE) exceeds a certain value for N consecutive periods, or when the rate of change of the drift metric exceeds a certain threshold.

[0070] Through the above technical solution, this application introduces a drift index calculation and monitoring mechanism in the real-time rolling forecasting process, which can promptly detect potential degradation in model performance or changes in data distribution. When the drift index meets preset trigger conditions, the system can automatically update the rolling window data of the training and validation sets and trigger the combined optimization algorithm to re-optimize the model hyperparameters. This adaptive re-optimization mechanism ensures that both the linear and nonlinear forecasting models can always learn the latest data patterns and adjust their training configuration to adapt to constantly changing power load characteristics and external environments. This significantly enhances the robustness and prediction accuracy of the hybrid model in long-term operation, effectively avoids model performance degradation caused by data drift, and thus guarantees the continuous high accuracy of real-time power load forecasting.

[0071] Example 6 Figure 2 This is a schematic diagram of the structure of a real-time power load forecasting system based on a hybrid model, provided in one embodiment of this application. (Refer to...) Figure 2 A real-time power load forecasting system based on a hybrid model includes: Processor 21 and memory 22; Processor 21 and memory 22 are connected via a communication bus: The processor 21 is used to call and execute the program stored in the memory 22; The memory 22 is used to store a program, which is used at least for a real-time power load forecasting method based on a hybrid model as described in the above embodiments.

[0072] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0073] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0074] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0075] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0076] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0077] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0078] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0079] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0080] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A real-time power load forecasting method based on a hybrid model, characterized in that, include: The current period's power load data is obtained and combined with historical period's power load data to form a power load sequence. At the same time, the weather data corresponding to the power load sequence is obtained. The power load sequence and the weather data are subjected to timestamp alignment, anomaly handling and normalization processing to obtain a standardized input sequence. Determine the order parameters of the linear prediction model and the set of hyperparameters to be optimized, consisting of the network structure parameters and training parameters of the nonlinear prediction model; The set of hyperparameters to be optimized is optimized by a combinatorial optimization algorithm, and the linear prediction model and the nonlinear prediction model are trained and updated using the optimized set of hyperparameters. The standardized input sequence is input into the nonlinear prediction model and the linear prediction model respectively. The nonlinear prediction model outputs a first prediction sequence, and the linear prediction model outputs a second prediction sequence. The prediction sequence is the power load prediction sequence for the next cycle in the current cycle. The first prediction sequence is compensated using the second prediction sequence as a trend term to obtain the fused prediction sequence; Obtain the fused prediction sequence for the current period from the output of the previous period. Calculate the residuals based on the fused prediction sequence of the previous period and the power load data of the current period, and construct a residual sequence. Determine the deviation term based on the residual sequence. Use the deviation term to perform online calibration on the fused prediction sequence of the current period, and output the calibrated load prediction result.

2. The method according to claim 1, characterized in that, Before inputting the standardized input sequence into the nonlinear prediction model and the linear prediction model respectively, the method further includes: Based on the standardized input sequence, a training set and a validation set are constructed to determine the set of hyperparameters to be optimized; the set of hyperparameters to be optimized includes at least: the order parameters of the linear prediction model, and the network structure parameters and training parameters of the nonlinear prediction model; The set of hyperparameters to be optimized is encoded to construct an initial population for a genetic algorithm, and the prediction error on the validation set is used as the fitness function. The initial population is iteratively updated through selection, crossover, and mutation to obtain a set of candidate optimal hyperparameters. Using the candidate set of preferred hyperparameters as the initial positions of particles in the particle swarm optimization algorithm, the particle velocity and position are updated based on the individual optimal solution and the global optimal solution, and the target hyperparameter combination is obtained iteratively. Based on the target hyperparameter combination, the nonlinear prediction model and the linear prediction model are updated with parameters and training configuration, and the updated models are used to output the first prediction sequence and the second prediction sequence, respectively.

3. The method according to claim 2, characterized in that, The method further includes: During real-time rolling prediction, the training parameters are adaptively adjusted based on the changes in the prediction residuals within the most recent preset window. The training parameters include at least the learning rate. Select a hidden layer size that matches the current residual level from the preset network size candidate set in order to maintain the adaptability of the nonlinear prediction model to load fluctuations at different time scales.

4. The method according to claim 1, characterized in that, Before performing normalization processing on the power load sequence and the weather data, the method further includes: A time feature sequence is generated based on the timestamp of the power load sequence; the time feature sequence includes time feature, weekday feature and holiday identifier; The time feature sequence and the weather data are used together as exogenous features to form a multivariate input sequence with the power load sequence.

5. The method according to claim 1, characterized in that, The first predicted sequence is compensated using the second predicted sequence as a trend term, including: The first prediction sequence and the second prediction sequence are timestamp aligned according to a preset prediction step size to obtain an aligned prediction sequence. Scale matching is performed on the second prediction sequence in the aligned prediction sequence so that the amplitude scale of the second prediction sequence is consistent with that of the first prediction sequence; The trend compensation parameters are determined based on the residual sequence within the most recent preset window; the most recent preset window is the time window between the previous period and the current period; the trend compensation parameters include at least a trend compensation coefficient and a bias term. Based on the trend compensation parameters, a trend compensation term is generated for the second prediction sequence, and the trend compensation term is superimposed on the first prediction sequence to obtain the fused prediction sequence.

6. The method according to claim 5, characterized in that, The trend compensation coefficient is a dynamic coefficient; the dynamic coefficient is updated according to the weather characteristics, time period characteristics or residual level of the current period to adjust the influence of the trend compensation term on the fused prediction sequence.

7. The method according to claim 6, characterized in that, The dynamic coefficients are updated based on the weather characteristics, time period characteristics, or residual level of the current period, including: Obtain the weather feature vector, time period feature vector, and residual sequence within the most recent preset window for the current period, and calculate the residual level index based on the residual sequence within the most recent preset window; the residual level index includes at least one or more of the residual mean, residual root mean square, or residual absolute mean. The weather feature vector, the time period feature vector, and the residual level index are input into the dynamic coefficient update function to obtain candidate values ​​for the dynamic coefficient. Boundary constraint processing is performed on the candidate values ​​to ensure that the dynamic coefficients are within a preset range; The dynamic coefficients are updated based on a smooth update strategy, which includes: weighting and fusing the dynamic coefficients of the previous period with the candidate values ​​at a preset update step size to obtain the dynamic coefficients of the current period. When the residual level index exceeds a preset threshold, the dynamic coefficient is reduced to decrease the impact of the trend compensation term; when the residual level index is below the preset threshold, the dynamic coefficient is increased to enhance the impact of the trend compensation term.

8. The method according to claim 2, characterized in that, The residuals are calculated and a residual sequence is constructed based on the fused prediction sequence of the previous period and the power load data of the current period. The deviation term is determined based on the residual sequence, including: The fused prediction sequence for the current period output from the previous period is timestamped with the power load data of the current period to obtain an aligned sample. Calculate the residuals based on the aligned samples and construct the residual sequence; The bias term is obtained by weighting the residual sequence according to the time decay weight. The deviation term is superimposed on the fusion prediction sequence of the current period to complete the online calibration.

9. The method according to claim 8, characterized in that, The method further includes: During real-time rolling forecasting, the drift index is calculated based on the residual sequence within the most recent preset window. When the drift metric meets the preset triggering condition, the rolling window data of the training set and validation set is updated, and the combined optimization algorithm is triggered to re-optimize the set of hyperparameters to be optimized, so as to update the training configuration of the linear prediction model and the nonlinear prediction model.

10. A real-time power load forecasting system based on a hybrid model, characterized in that, include: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute the real-time power load forecasting method based on a hybrid model as described in any one of claims 1-9.

Images