Power load prediction method

Abstract

The invention discloses a power load prediction method, and the method comprises the steps: firstly solving an extreme event data sparsity problem through a generative adversarial network, and constructing an event time sequence library through a time sequence anomaly detection algorithm; then analyzing the causal relationship between the event and the load by applying a causal discovery algorithm, and converting prediction output into probability distribution by adopting a Bayesian neural network to quantify uncertainty; constructing a prediction model triggered by an event, and generating a multi-time scale probability prediction interval; and finally, generating a multi-scene prediction result through Monte Carlo simulation, quantifying the system recovery capability in combination with a toughness index, and integrating the system recovery capability to a decision support system to generate a risk response scheme. According to the method, the accuracy and robustness of load prediction under the extreme climate are remarkably improved, full-chain risk insight from early warning to recovery is realized, and prospective decision support is provided for safe operation of a power system.

Description

Technical Field

[0001] This invention relates to the field of power technology, and in particular to a method for predicting power load. Background Technology

[0002] Traditional power load forecasting methods primarily rely on time-series correlation analysis between historical load data and conventional meteorological indicators (such as temperature and humidity), employing statistical models or machine learning algorithms for trend extrapolation. These methods exhibit significant limitations when facing increasingly frequent extreme weather events. Firstly, the scarcity of extreme event samples leads to a highly uneven distribution of training data, making it difficult for models to accurately learn the complex nonlinear relationship between extreme events and load fluctuations, resulting in predictive biases. Secondly, traditional methods are often based on the assumption of "single-point events," failing to effectively capture the cumulative impact on the power system from causally related chains of events (i.e., cascading effects), such as "drought-high temperature-wildfire." Furthermore, models typically output deterministic point predictions, failing to quantify the uncertainty of the prediction results. This lack of probabilistic basis for risk assessment makes it difficult to support grid dispatch decisions with high reliability requirements.

[0003] While existing research attempts to introduce more complex deep learning models to improve prediction accuracy, several key challenges remain. First, most models employ deterministic architectures, failing to systematically separate and quantify sources of uncertainty such as data noise, model approximation errors, and event randomness, resulting in coarse risk measurements. Second, the prediction process is disconnected from system resilience assessment, failing to directly translate load forecasting results into quantitative support for resilience indicators such as system recovery time and loss severity, thus limiting its practicality in emergency response planning. More importantly, existing methods typically employ static modeling approaches, lacking the ability to adapt to the dynamic evolution of extreme events. They cannot adjust prediction strategies in real time according to the development stages of the event chain (e.g., germination, outbreak, and decline), leading to a significant decline in prediction performance during rapid evolution of sudden events. These shortcomings constrain the power system's proactive defense and rapid recovery capabilities in response to extreme weather risks. Summary of the Invention

[0004] This application provides a power load forecasting method that significantly improves the accuracy of load forecasting and grid resilience under extreme weather conditions, providing decision support for the safe operation of the power system.

[0005] This application provides a method for predicting electricity load, including:

[0006] S1 uses generative adversarial networks to augment data by collecting real-time and historical data from heterogeneous data sources, addressing the problem of data sparsity in extreme events.

[0007] S2 uses a time-series-based anomaly detection algorithm to scan climate data in real time, identify the occurrence points, intensity, and duration of extreme events, and classify the detected events to form an event time series library;

[0008] S3 applies a causal discovery algorithm to analyze the causal relationship between events and load fluctuations, while identifying the main sources of forecast uncertainty;

[0009] S4 uses a Bayesian neural network to transform the output of the load forecasting network from a single-point estimate into a probability distribution in order to quantify the uncertainty introduced by the model and data;

[0010] S5. Construct an event-triggered time series forecasting model. Using the aforementioned event time series library as dynamic input and combining the uncertainty quantification results, achieve load forecasting at multiple time scales and generate probabilistic forecasting intervals.

[0011] S6, based on probabilistic prediction intervals, generates multi-scenario prediction results through Monte Carlo sampling, and combines resilience index to quantitatively evaluate the load performance and recovery capability of electricity under extreme events;

[0012] S7 integrates load forecasting, uncertainty quantification, and resilience assessment results into decision support, and generates risk response plans through a rules engine.

[0013] Preferably, the data augmentation using a generative adversarial network (GAN) specifically includes: the generator of the GAN learns the potential distribution patterns of real historical extreme event data, receives random noise vectors, and outputs simulated extreme climate data sequences; the discriminator of the GAN distinguishes between generated data and real extreme event data, and outputs the probability that the data is real data; through iterative adversarial training between the generator and the discriminator until Nash equilibrium is reached, the generated synthetic data is merged with real historical data, and the merged dataset is standardized. 3. The power load forecasting method as described in claim 1, characterized in that the use of a time series-based anomaly detection algorithm specifically employs an autoencoder based on a long short-term memory network; the autoencoder is pre-trained on historical normal climate data, and anomaly scores are obtained by calculating the reconstruction error between the input climate subsequence and the reconstructed sequence; a dynamic detection threshold is set based on the reconstruction error sequence, and when the anomaly score continuously exceeds the threshold, an extreme event is determined to have started, and quantitative feature indicators of the event are extracted for classification and labeling, and stored in an event time series database.

[0014] Preferably, the applied causal discovery algorithm is the PC algorithm; the PC algorithm infers the network structure of potential causal relationships between variables by performing conditional independence tests, forming a causal directed acyclic graph; based on the causal graph, structural equation modeling or regression analysis is used to quantify the effect strength of the causal path, and the sources of uncertainty are classified and characterized.

[0015] Preferably, the quantitative assessment of power load performance and recovery capability under extreme events specifically includes: using the load prediction probability distribution to independently sample each future time point; repeating this process a certain number of times to generate a set of scenarios representing possible future load evolution paths, with each scenario constituting a complete load time series; calculating the resilience index for each generated load scenario; statistically aggregating the calculation results of all scenarios to analyze the distribution characteristics of recovery time, loss value, and resilience score, thereby deriving a probabilistic assessment conclusion on resilience.

[0016] Preferably, step S2, which identifies the occurrence point, intensity, and duration of extreme events, further includes: S21, using a sequence pattern recognition algorithm based on heterogeneous data sources to detect and classify event chains in real time, and mining the association between their key nodes and historical load fluctuation patterns to construct an event chain-load correlation feature library; S22, using a causal discovery algorithm to construct a causal graph network containing event chain nodes and load factors, and extracting key impact paths to quantify their potential impact on the load; S23, using stochastic processes to simulate the evolution path of the event chain, and dynamically adjusting the feature focusing strategy and uncertainty parameters of the load forecasting model based on path risk; S24, integrating the uncertainty of the event chain and the model to generate a load forecasting result with a probabilistic resilience index.

[0017] Preferably, the step of extracting key impact paths to quantify their potential impact on load includes: using a constraint-based causal discovery algorithm to analyze multivariate time series containing climate variables, event chain nodes, infrastructure status, and load factors; performing conditional independence tests to eliminate spurious associations between variables and inferring the structure of a directed acyclic graph describing the causal direction between variables; based on the obtained causal graph, using structural equation modeling or a do-calculus framework to estimate causal effects on directed edges from event chain nodes to load factors, quantifying the average impact of event chain nodes on load values ​​under controlled confounding factors; extracting all directed paths from the initial event chain trigger point to load variables from the causal graph; calculating the strength of each path, characterized by the product or minimum value of the causal effects of each link in the path; and calculating the risk value of each path on the load by combining the path strength with the probability of event chain occurrence, thereby achieving a probabilistic assessment of the potential impact on load.

[0018] Preferably, the method further includes: the path strength calculation formula is: for a causal path P: A→B→C, its path strength can be defined as: CE represents the estimated causal effect of each edge; the risk value of each path to the load is assessed by combining path strength and event probability; its calculation formula is: P(A) is the probability of the path starting event A occurring, Strength(P) is the strength of the path, and Impact is the severity coefficient of the load impact.

[0019] Preferably, the real-time detection and classification of event chains based on heterogeneous data sources using sequence pattern recognition algorithms further includes: establishing a collaborative prediction mechanism by constructing a distributed information sharing network for the microgrid cluster, integrating climate and load data from multiple nodes; designing a prediction coordination algorithm for the microgrid cluster, and improving the overall prediction accuracy of the microgrid cluster through dynamic weight allocation; automatically assigning a node with load-sharing capabilities that meets the requirements of a node when an extreme climate event chain is detected by a node in the microgrid cluster based on the event chain detection results; and constructing a multi-dimensional resilience assessment system for the microgrid cluster, and monitoring the mutual support capabilities of nodes and the reliability of predictions in real time.

[0020] Preferably, the load capacity specifically includes: a continuous operation event chain detection algorithm for each node in the microgrid group; when a node detects an extreme weather event chain and its risk index exceeds a preset threshold, it broadcasts an event alarm to the group; after receiving the alarm, it verifies the event chain type, impact range, and node load status to determine whether to initiate load diversion; it collects the operating parameters of each node in the group in real time, including the current load factor, maximum carrying capacity, backup power status, line transmission margin, and local event risk level; it uses a multi-index fusion method to calculate the real-time load capacity and diversion priority of each node; it establishes a load diversion optimization model with the goal of minimizing the total operating risk, with constraints including node capacity limits, power balance equations, and network security constraints; it solves the optimal load transfer scheme between nodes through a distributed optimization algorithm; it converts the optimal diversion scheme into control commands for execution to ensure a smooth load transition; it monitors the operating status of each node after diversion in real time to verify whether it is in a safe and stable range; wherein, the calculation formula for the real-time load capacity of each node is: , This represents the amount of additional load that node i can currently accept. This represents the maximum allowable load for this node. For its current actual load, This is a safety margin calculated based on the local event risk level.

[0021] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0022] By integrating multi-source data with a dynamic event perception mechanism, the accuracy and robustness of power load forecasting under extreme weather conditions are significantly improved. Uncertainty quantification and causal analysis enhance the interpretability and decision reliability of the model. Furthermore, a resilience assessment system provides a full-chain risk insight for the safe operation of the power grid, from early warning to recovery, ultimately achieving forward-looking and adaptive optimization of power resource scheduling.

[0023] By introducing dynamic event chain perception and causal reasoning mechanisms, the accuracy and foresight of power load fluctuation prediction under extreme weather events are significantly improved; through stochastic process simulation and uncertainty fusion, the probabilistic quantification of cascading failure risks is realized, enhancing the robustness and decision reliability of the model; the finally generated resilience index provides a full-chain risk insight for the safe operation of the power grid from early warning to recovery, supporting adaptive resource scheduling and resilience optimization.

[0024] By introducing a distributed load forecasting mechanism based on microgrid group collaboration, the accuracy and overall resilience of power load forecasting under extreme weather conditions are significantly improved. This scheme achieves dynamic allocation and fairness assurance of forecasting resources through information sharing and game theory optimization. A climate event-driven model switching strategy enhances the ability to respond quickly to emergencies, while a multi-dimensional resilience assessment system provides forward-looking decision support for grid safety operation through real-time monitoring and mutual assistance analysis. Ultimately, this effectively reduces forecast uncertainty, optimizes resource scheduling, and strengthens the adaptive and risk-resistant capabilities of the power grid. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a power load forecasting method according to an embodiment of the present invention. Detailed Implementation

[0026] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] Example 1: Figure 1 This is a flowchart illustrating a power load forecasting method according to an embodiment of the present invention.

[0029] like Figure 1 As shown, a power load forecasting method includes the following steps:

[0030] S1 uses generative adversarial networks to augment data by collecting real-time and historical data from heterogeneous data sources, addressing the problem of data sparsity in extreme events.

[0031] Among them, heterogeneous data sources include climate data (such as satellite remote sensing temperature, precipitation, and wind speed), power load data (such as regional electricity consumption records), power grid operation data (such as generator output), socio-economic data (such as GDP and policy changes), and event reports (such as disaster early warnings).

[0032] Specifically, historical and real-time data are synchronously acquired from heterogeneous data sources, and all data sequences are unified to the same time granularity, for example, all are converted into hourly data points.

[0033] Based on historical records, the periods in which extreme events occurred are marked in existing climate data sequences; for example, the date range corresponding to a sustained heat wave is marked. The proportion of the marked extreme event data points in the overall dataset is statistically analyzed to confirm that it constitutes a sparse dataset.

[0034] The generator is a neural network that aims to learn the underlying distribution patterns of extreme event data from real history, such as temperature and wind speed sequences during heat waves. It receives a random noise vector and attempts to output a convincingly simulated extreme weather data sequence. The discriminator is another neural network that aims to distinguish the data generated by the generator from real extreme event data. It receives a data sequence and outputs a scalar between 0 and 1, representing the probability that the sequence is real data. The two networks compete and co-evolve. The generator attempts to generate realistic data to fool the discriminator, while the discriminator improves its ability to distinguish between real and generated data. This process iterates until a Nash equilibrium is reached, where the discriminator can no longer effectively distinguish between real and generated data. At this point, the generator can produce high-quality synthetic extreme event data.

[0035] The large amount of synthetic extreme event data generated by the generator is merged with real historical data to form a balanced augmented dataset. The augmented dataset is then standardized to eliminate the influence of different physical dimensions and to prepare for subsequent model training.

[0036] S2 uses a time-series-based anomaly detection algorithm to scan climate data in real time, identify the occurrence points, intensity, and duration of extreme events, and classify the detected events to form an event time-series database.

[0037] Specifically, the climate data from heterogeneous data sources in step S1 are divided into fixed time windows for continuous analysis.

[0038] An autoencoder based on a long short-term memory network is used. The model is pre-trained on a large amount of normal climate data to learn how to reconstruct normal climate sequence patterns with high accuracy.

[0039] The climate data sequence, after enhancement and standardization in step S1, is divided into windows of fixed time length as the basic analysis unit for event detection.

[0040] An autoencoder model based on a long short-term memory network architecture is employed. This model has been pre-trained on historical normal climate data, and its optimization objective is to minimize the difference between the input sequence and the reconstructed sequence. For each climate subsequence within a sliding window, the autoencoder performs encoding and decoding operations to generate the corresponding reconstructed sequence. The anomaly score for that time window is obtained by calculating the reconstruction error between the original sequence and the reconstructed sequence.

[0041] Based on the recent statistical characteristics of the reconstructed error sequence, a time-varying dynamic detection threshold is set. When consecutive observations in the abnormal score sequence exceed this threshold, it is determined as the start signal of an extreme event. The termination point of the event is determined as the time point when the abnormal score continuously falls back below the threshold.

[0042] For each identified event, quantitative characteristic indicators are extracted from its corresponding raw climate data, including average intensity, peak intensity, and cumulative intensity over the event's duration. Based on a pre-defined classification rule system based on physical characteristics and impact range, the events are standardized and labeled according to their type and level.

[0043] The event's metadata (start and end times, duration), quantitative feature vectors, and classification labels are integrated into a structured record and stored in a time-indexed event time series database, providing support for subsequent causal analysis and predictive modeling.

[0044] S3 applies a causal discovery algorithm to analyze the causal relationship between events and load fluctuations, while identifying the main sources of forecast uncertainty.

[0045] Specifically, a constraint-based causal discovery algorithm, such as the PC algorithm, is employed to analyze multivariate time series data encompassing climate variables, identified event characteristics, load data, and other covariates. This algorithm performs a series of conditional independence tests to progressively eliminate redundant correlations between variables, inferring the network structure describing the potential causal relationships between them—that is, a causal directed acyclic graph.

[0046] Based on the obtained causal graph structure, for the identified significant causal paths, such as from "abnormal temperature" to "industrial electricity load", methods such as structural equation modeling or regression analysis are used to quantify the magnitude and direction of their causal effects.

[0047] Based on causal analysis results and domain knowledge, the root causes of uncertainty in prediction models are categorized. The main categories include: data noise from the data acquisition and processing process, model approximation errors introduced by the model's simplification of complex real-world relationships, and inherent uncertainties of events stemming from the inherent randomness of climate and the unpredictability of event evolution paths.

[0048] Preliminary characterizations are performed on the identified sources of uncertainty. For example, data noise can be characterized by sensor accuracy metrics or the statistical distribution of historical errors; model error can be estimated by performance deviation on the validation set; and event randomness can be approximated by the dispersion of outcomes of similar historical events.

[0049] S4 employs a Bayesian neural network to transform the output of the load forecasting network from a single-point estimate into a probability distribution, in order to quantify the uncertainty introduced by the model and data.

[0050] Specifically, the weight parameters in traditional deterministic neural networks are redefined as probability distributions. Each weight is no longer a deterministic scalar, but a random variable that follows a prior distribution (such as a Gaussian distribution).

[0051] Variational inference is used for model training. The goal is to find a computationally friendly approximate distribution (such as a Gaussian distribution) to approximate the true, but difficult-to-compute, theoretical weight posterior distribution.

[0052] For a given input, the predicted value is calculated through forward propagation. Since the model weights are distributed, the set of weights sampled in each forward propagation is different, resulting in different predicted values. By sampling multiple times, an empirical distribution of the load forecast can be obtained.

[0053] The total forecast uncertainty is decomposed into cognitive uncertainty (originating from the uncertainty of model parameters, which decreases as the amount of data increases) and random uncertainty (originating from the inherent noise of the data, which is an inherent and unavoidable part).

[0054] S5. Construct an event-triggered time series forecasting model. Using the aforementioned event time series library as dynamic input and combining the uncertainty quantification results, achieve load forecasting at multiple time scales and generate probabilistic forecast intervals.

[0055] Specifically, a Bayesian Long Short-Term Memory (LSTM) network is used as the core prediction engine. The network weights are based on a probability distribution, and its forward propagation process is essentially a stochastic process. The network output layer is designed with a multi-branch structure, corresponding to short-term (hourly), medium-term (daily), and long-term (monthly) load prediction targets, respectively.

[0056] Event feature vectors (such as type, intensity, and duration) extracted from an event time series database are input into the network as additional dynamic context information. The model employs an event-triggered gating mechanism, transforming the event feature vectors and using them as modulation factors applied to the forget gate or input gate of the LSTM unit. When a strong event signal is input, this mechanism automatically enhances the network's memory and attention to recent mutation patterns, while weakening long-term historical dependencies unrelated to the current event.

[0057] For a given input sequence, forward propagation is performed through multiple Monte Carlo sampling operations. Each sampling sample represents a set of network weights, resulting in a load prediction sequence for each time scale. The predicted values ​​for all future time points across all sampling counts are aggregated to form an empirical distribution of the load at that time point. Based on this distribution, the prediction mean and the prediction interval at a specified confidence level (e.g., 95%) are calculated. The training objective of the model is to minimize the joint loss between the predicted results and the actual values ​​across all time scales. The loss function must consider the accuracy of the predictions at each scale and the degree of uncertainty calibration.

[0058] S6, based on probabilistic prediction intervals, generates multi-scenario prediction results through Monte Carlo sampling, and combines resilience index to quantitatively assess the load performance and recovery capability of electricity under extreme events.

[0059] The resilience index includes recovery time estimates (such as the number of hours required for the load to return to normal), loss quantification (such as power gaps caused by forecasting errors), and resilience scores (based on historical event recovery capabilities).

[0060] Specifically, using the load forecast probability distribution (mean and variance) obtained in step S5, each future time point is sampled independently. By repeating this process a large number of times, a set of scenarios representing possible future load evolution paths is generated, with each scenario constituting a complete load time series.

[0061] For each generated load scenario, its resilience index is calculated. Recovery time estimation identifies the starting point when the load value first falls below a preset normal threshold in the scenario sequence, and the ending point when it first recovers and remains above that threshold; the difference between these two points represents the estimated recovery duration for that scenario. Loss quantification involves integrating the periods in the scenario sequence where the load value is below the normal threshold, calculating the "deficit" area enclosed by this area and the normal load level, thus quantifying the cumulative power supply gap in that scenario. Resilience scoring compares the recovery time of this scenario with typical recovery times in similar historical extreme events, obtaining a relative score through normalization to reflect recovery efficiency compared to historical performance.

[0062] The calculation results for all scenarios are statistically aggregated. The distribution characteristics (such as mean and quantile) of recovery time, loss value and resilience score are analyzed to draw probabilistic assessment conclusions of resilience, such as "there is a 90% probability that the recovery time will be within 12 hours".

[0063] S7 integrates load forecasting, uncertainty quantification, and resilience assessment results into decision support, and generates risk response plans through a rules engine.

[0064] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0065] By integrating multi-source data with a dynamic event perception mechanism, the accuracy and robustness of power load forecasting under extreme weather conditions are significantly improved. Uncertainty quantification and causal analysis enhance the interpretability and decision reliability of the model. Furthermore, a resilience assessment system provides a full-chain risk insight for the safe operation of the power grid, from early warning to recovery, ultimately achieving forward-looking and adaptive optimization of power resource scheduling.

[0066] Example 2: In Example 1, by acquiring climate, load, and electricity prices of the target region, and inputting them into a preset power load prediction model, time-series features at different time steps are extracted and fused to output power load prediction data of various time lengths. However, this method may not be able to effectively capture the dynamic correlations and chain reactions between events when dealing with extreme weather events. For example, when a single extreme event (such as a heat wave or hurricane) triggers a series of secondary events (such as infrastructure failures or load surges), relying solely on a unified feature fusion network may not accurately quantify the uncertainty propagation path, resulting in large prediction biases and weak generalization ability. Since the evolution speed, impact range, and load fluctuation patterns of different event chains vary significantly, using a static model structure may introduce training errors when facing multi-scale prediction needs in complex scenarios, making it difficult to adapt to causal relationships in ever-changing real-world scenarios. To improve prediction accuracy and resilience, it is necessary to simultaneously consider the dynamic perception of event chains, causal reasoning, and uncertainty fusion for further optimization and improvement.

[0067] In some embodiments, in addition to identifying the location, intensity, and duration of the extreme event, step S2 further includes:

[0068] S21. Based on heterogeneous data sources, a sequence pattern recognition algorithm is used to detect and classify event chains in real time, and the key nodes are associated with historical load fluctuation patterns to construct an event chain-load correlation feature library.

[0069] Specifically, heterogeneous data sources are time-aligned and cleaned to unify them to the same time granularity, forming a multivariate time series dataset.

[0070] Sequence pattern recognition algorithms (such as dynamic time warping or neural processes) are used to analyze multivariate sequences and identify continuous events that conform to specific evolutionary patterns, i.e., the trigger points of event chains. Based on the evolution speed of event chains (such as slowly accumulating drought or rapidly erupting wildfires) and the spatial impact range, they are classified (such as slow-onset chains and fast-onset chains), and the type, start time, and key node sequence of event chains are recorded (the core event points or state mutation points that mark state transitions and directly trigger specific pattern fluctuations in power load during the evolution of event chains).

[0071] The identified event chain timelines are rigorously aligned with historical load data. For each event chain, the specific fluctuation patterns in load data before and after its key nodes (such as "wildfire occurrence" or "power outage") are analyzed (e.g., industrial electricity consumption drops sharply after "power outage"). Association rule mining techniques are used to quantify the correlation strength between event chain nodes and load patterns.

[0072] The discovered association rules (including event chain types, key nodes, load fluctuation patterns, and association strength indicators) are integrated into a structured feature library, which serves as the event-load association knowledge base for the prediction model.

[0073] S22, a causal graph network containing event chain nodes and load factors is constructed using a causal discovery algorithm, and key impact paths are extracted to quantify their potential impact on the load.

[0074] Specifically, constraint-based causal discovery algorithms (such as PC or FCI algorithms) are employed to analyze multivariate time series data including climate variables, event chain nodes, infrastructure status, and load factors. Conditional independence tests are performed selectively to eliminate spurious associations between variables, and a directed acyclic graph structure describing the causal direction among the variables is inferred.

[0075] Based on the obtained causal graph, structural equation modeling or do-calculus framework is used to estimate causal effects for directed edges from event chain nodes to load factors, quantifying the average influence of event chain nodes on load values ​​under the control of confounding factors.

[0076] Extract all directed paths from the initial event chain trigger point (e.g., "drought index exceeding the limit") to the load variable (e.g., "regional total load") from the causal graph. For each path, calculate its strength, which is characterized by the product or minimum of the causal effects of each link in the path.

[0077] By combining path strength and the probability of event chain occurrence, the risk value of each path to the load is calculated, thereby achieving a probabilistic assessment of the potential impact on the load.

[0078] The path strength calculation formula is as follows: For a causal path P: A→B→C, its path strength can be defined as:

[0079]

[0080] CE represents the estimated causal effect of each edge (such as regression coefficient or ATE). The minimum value method follows the "barrel principle" and emphasizes the reliability of the path; the multiplication method considers the cumulative effect.

[0081] The risk value posed by each path to the load is assessed by combining path strength and the probability of event occurrence. The calculation formula is as follows: P(A) is the probability of path origin event A occurring, Strength(P) is the strength of the path, and Impact is the severity coefficient of the load impact (which can be set according to historical data).

[0082] S23 utilizes stochastic processes to simulate the evolution path of event chains and dynamically adjusts the feature-focusing strategy and uncertainty parameters of the load forecasting model based on path risk.

[0083] Specifically, the causal directed acyclic graph constructed in stage S22 is transformed into a continuous-time Markov chain or a dynamic Bayesian network. In this stochastic process model, the state of an event chain node (e.g., "not occurred" or "occurred") is defined as a random variable, and its state transition probability is determined by the edge weights (causal strength) in the causal graph. Through Monte Carlo simulation, a large number of event chain evolution paths are generated, covering scenarios including event chain interruption, continuous deterioration, or the emergence of new events.

[0084] Based on the generated event chain path, the load forecasting model (such as a Bayesian long short-term memory network) is dynamically adjusted. When the path indicates that the event chain is in an early or low-risk state, the model strengthens the fusion of long-term features to predict trend load changes; when the path indicates that the event chain has entered a critical node or a high-risk state, the model switches to a short-term feature focusing mode and increases uncertainty parameters (such as increasing the prior distribution variance) to output a load probability distribution with a wider confidence interval.

[0085] Adaptive prediction is performed on each event chain path, and the results are aggregated through Monte Carlo simulation to generate the probability distribution of load at each future time point, representing the range of load values ​​and the probability of occurrence under different scenarios.

[0086] S24 integrates the event chain and model uncertainty to generate load forecast results with a probabilistic resilience index.

[0087] Specifically, based on the event chain path and load prediction distribution generated in stage S23, the total prediction uncertainty is decomposed using the full variance decomposition method, and the contribution of event chain stochastic evolution and model cognitive limitations to uncertainty is quantified.

[0088] To define a resilience index specific to event chains, firstly, based on the simulated path of the event chain, the probability distribution of the time required for the load to return to normal levels is predicted, and the expected value and quantile of the recovery time are calculated. Secondly, by combining the load prediction deviation, the expected recovery time, and the event chain strength, a resilience scoring function is constructed to evaluate the overall adaptability and recovery capability of the power system.

[0089] The load forecast, the comprehensive uncertainty range, and the resilience index are integrated into a structured report, which clarifies the core indicators, uncertainty range, and resilience level of load forecast under different event chain scenarios.

[0090] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0091] By introducing dynamic event chain perception and causal reasoning mechanisms, the accuracy and foresight of power load fluctuation prediction under extreme weather events are significantly improved; through stochastic process simulation and uncertainty fusion, the probabilistic quantification of cascading failure risks is realized, enhancing the robustness and decision reliability of the model; the finally generated resilience index provides a full-chain risk insight for the safe operation of the power grid from early warning to recovery, supporting adaptive resource scheduling and resilience optimization.

[0092] Example 3: In Example 2, centralized uncertainty quantification and resilience assessment for load forecasting under extreme weather conditions were achieved through dynamic event chain perception, causal graph network construction, and stochastic process simulation, effectively improving forecast accuracy and robustness. However, this method does not fully consider the synergistic effects and mutual assistance potential of distributed microgrid environments. When extreme weather events simultaneously affect multiple regions, centralized forecasting may be biased due to information transmission delays or missing local data. Furthermore, the lack of dynamic optimization of microgrid group behavior limits the generalization ability of the forecasting model in complex scenarios. Since load fluctuations among distributed nodes exhibit spatiotemporal correlation and resource complementarity, a unified centralized model is insufficient to adapt to the uncertainties brought about by multi-node interactions. To further improve the response speed and forecast reliability for wide-area extreme events, a microgrid group collaboration mechanism needs to be introduced to achieve optimization and improvement from centralized reasoning to distributed intelligent decision-making.

[0093] In some embodiments, step S21, which involves using a sequence pattern recognition algorithm based on heterogeneous data sources to detect and classify event chains in real time, further includes:

[0094] S211 establishes a collaborative forecasting mechanism by constructing a distributed information sharing network for microgrid clusters, integrating climate and load data from multiple nodes.

[0095] Specifically, based on peer-to-peer communication protocols or a lightweight blockchain architecture, secure communication links are established between microgrid nodes, forming a decentralized information-sharing network. The network topology supports dynamic adjustments based on node operating status to ensure robustness and efficiency of information transmission. Each microgrid node maintains a local forecasting model. A collaborative mechanism stipulates that before performing load forecasting, a node requests relevant feature data (such as historical load patterns and recent climate anomalies) from its neighboring nodes. By fusing local features with those of neighboring nodes, an enhanced input feature set is constructed and used as input to the local forecasting model.

[0096] S212, Design a prediction coordination algorithm for microgrid groups, and improve the overall prediction accuracy of microgrid groups through dynamic weight allocation.

[0097] Specifically, the microgrid cluster forecasting coordination problem is modeled as a cooperative game theory model. In this framework, each microgrid is considered a participant, and all participants form a cooperative alliance. Each participant achieves higher accuracy than independent forecasting through information sharing and collaborative forecasting. The policy space is defined as the degree of contribution each participant makes to the local forecasting results and the weight of the influence of neighboring node forecasts.

[0098] Define a utility function for each participant to quantify their net gain from participating in the consortium. The utility function is determined by the gains from improved prediction accuracy and the costs of collaboration (including communication overhead and computational resource consumption). The optimization objective is to find the strategy combination that maximizes the overall utility of the consortium, or to reach a Nash equilibrium.

[0099] Based on a game theory model, a dynamic weight allocation algorithm is designed. This algorithm dynamically calculates the weight of each node in the group prediction based on its real-time prediction reliability, local event risk level, and marginal contribution to the alliance. Nodes with high prediction reliability and located on high-risk paths in the event chain are assigned higher weights. Through iterative optimization, the weight allocation gradually converges to a stable state.

[0100] Convergence conditions are set, such as the norm of the weight vector change being less than a preset threshold in consecutive iterations, or the growth rate of the overall utility of the coalition being lower than a specific threshold. Finally, based on the converged weight vector, the group prediction results are weighted and fused to output a collaborative prediction value.

[0101] S213, based on the event chain detection results, when a node in the microgrid group detects an extreme weather event chain, it automatically assigns a node in the microgrid group that meets the node's load diversion capability.

[0102] Specifically, each node in the microgrid cluster continuously runs the event chain detection algorithm (S21 in Example 2). When a node detects an extreme weather event and its risk index exceeds a preset threshold, it broadcasts an event alarm to the cluster. Upon receiving the alarm, it verifies the event chain type, the scope of impact, and the node load status to determine whether to initiate load diversion.

[0103] The system collects real-time operating parameters of each node within the cluster, including current load factor, maximum carrying capacity, backup power status, line transmission margin, and local event risk level. A multi-indicator fusion method is used to calculate the real-time load capacity and load shedding priority of each node.

[0104] With the goal of minimizing overall operational risk, a load sharing optimization model is established, with constraints including node capacity limitations, power balance equations, and network security constraints. The optimal load transfer scheme between nodes is obtained by solving the model using a distributed optimization algorithm (such as the alternating direction multiplier method).

[0105] The optimal load splitting scheme is translated into control commands and executed to ensure a smooth load transition. The operating status of each node after load splitting is monitored in real time to verify whether it is within a safe and stable range.

[0106] The formula for calculating the real-time load capacity that each node can accept is as follows:

[0107]

[0108] This represents the amount of additional load that node i can currently accept. This represents the maximum allowable load for this node. This is its current actual load. This is a safety margin calculated based on the local event risk level.

[0109] S214 constructs a multi-dimensional resilience assessment system for microgrid clusters, which monitors the mutual support capabilities of nodes and predicts reliability in real time.

[0110] Specifically, a multi-dimensional resilience index system is defined, encompassing inter-node load sharing capability, predicted reliability propagation efficiency, and recovery coordination. The sharing capability characterizes the load transfer potential between nodes; predicted reliability quantifies the stability of prediction results; and recovery coordination assesses the coordination of node recovery behavior after a failure.

[0111] Through a distributed sensor network, real-time data such as load, prediction error, event chain impact status, and resource sharing records of each node are continuously collected. Based on the data within the sliding time window, the values ​​of various resilience indicators are updated periodically. For example, mutual assistance capability is calculated based on the frequency of load transfer requests and responses between nodes; prediction reliability is based on mean absolute error statistics; and recovery coordination is obtained through analysis of the time deviation of historical recovery events.

[0112] A weighted fusion method is used to synthesize the indicators from each dimension into an overall resilience score, with the weights reflecting the relative importance of each indicator in resilience. The score result is represented numerically to indicate the current resilience level.

[0113] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0114] By introducing a distributed load forecasting mechanism based on microgrid group collaboration, the accuracy and overall resilience of power load forecasting under extreme weather conditions are significantly improved. This scheme achieves dynamic allocation and fairness assurance of forecasting resources through information sharing and game theory optimization. A climate event-driven model switching strategy enhances the ability to respond quickly to emergencies, while a multi-dimensional resilience assessment system provides forward-looking decision support for grid safety operation through real-time monitoring and mutual assistance analysis. Ultimately, this effectively reduces forecast uncertainty, optimizes resource scheduling, and strengthens the adaptive and risk-resistant capabilities of the power grid.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting electricity load, characterized in that, include: S1 uses generative adversarial networks to augment data by collecting real-time and historical data from heterogeneous data sources, addressing the problem of data sparsity in extreme events. S2 uses a time-series-based anomaly detection algorithm to scan climate data in real time, identify the occurrence points, intensity, and duration of extreme events, and classify the detected events to form an event time series library; S3 applies a causal discovery algorithm to analyze the causal relationship between events and load fluctuations, while identifying the main sources of forecast uncertainty; S4 uses a Bayesian neural network to transform the output of the load forecasting network from a single-point estimate into a probability distribution in order to quantify the uncertainty introduced by the model and data; S5. Construct an event-triggered time series forecasting model. Using the aforementioned event time series library as dynamic input and combining the uncertainty quantification results, achieve load forecasting at multiple time scales and generate probabilistic forecasting intervals. S6, based on probabilistic prediction intervals, generates multi-scenario prediction results through Monte Carlo sampling, and combines resilience index to quantitatively evaluate the load performance and recovery capability of electricity under extreme events; S7 integrates load forecasting, uncertainty quantification, and resilience assessment results into decision support, and generates risk response plans through a rules engine.

2. The power load forecasting method as described in claim 1, characterized in that, The data augmentation using generative adversarial networks specifically includes: the generator of the generative adversarial network learns the potential distribution patterns of real historical extreme event data, receives random noise vectors and outputs simulated extreme climate data sequences; the discriminator of the generative adversarial network distinguishes between generated data and real extreme event data, and outputs the probability that the data is real data; through iterative adversarial training between the generator and the discriminator until Nash equilibrium is reached, the generated synthetic data is merged with real historical data, and the merged dataset is standardized.

3. The power load forecasting method as described in claim 1, characterized in that, The time-series-based anomaly detection algorithm specifically employs an autoencoder based on a long short-term memory network. The autoencoder is pre-trained on historical normal climate data, and anomaly scores are obtained by calculating the reconstruction error between the input climate subsequence and the reconstructed sequence. A dynamic detection threshold is set based on the reconstruction error sequence. When the anomaly score continuously exceeds the threshold, an extreme event is determined to have started, and quantitative feature indicators of the event are extracted for classification and labeling, and stored in the event time series database.

4. The power load forecasting method as described in claim 1, characterized in that, The proposed causal discovery algorithm is the PC algorithm. The PC algorithm infers the network structure of potential causal relationships between variables by performing conditional independence tests, forming a causal directed acyclic graph. Based on the causal graph, structural equation modeling or regression analysis is used to quantify the effect strength of the causal path and classify and characterize the sources of uncertainty.

5. The power load forecasting method as described in claim 1, characterized in that, The quantitative assessment of power load performance and recovery capability under extreme events specifically includes: using the load prediction probability distribution to independently sample each future time point; repeating this process a certain number of times to generate a set of scenarios representing possible future load evolution paths, with each scenario constituting a complete load time series; calculating the resilience index for each generated load scenario; statistically aggregating the calculation results of all scenarios to analyze the distribution characteristics of recovery time, loss value, and resilience score, thereby deriving a probabilistic assessment conclusion on resilience.

6. The power load forecasting method as described in claim 1, characterized in that, The step S2, which identifies the occurrence point, intensity, and duration of extreme events, further includes: S21, using a sequence pattern recognition algorithm based on heterogeneous data sources to detect and classify event chains in real time, and mining the association between key nodes and historical load fluctuation patterns to construct an event chain-load correlation feature library; S22, using a causal discovery algorithm to construct a causal graph network containing event chain nodes and load factors, and extracting key impact paths to quantify their potential impact on the load; S23, using stochastic processes to simulate the evolution path of the event chain, and dynamically adjusting the feature focusing strategy and uncertainty parameters of the load forecasting model based on path risk; S24, integrating the uncertainty of the event chain and the model to generate a load forecasting result with a probabilistic resilience index.

7. The power load forecasting method as described in claim 6, characterized in that, The process of extracting key impact paths to quantify their potential impact on load includes: analyzing multivariate time series data containing climate variables, event chain nodes, infrastructure status, and load factors using a constraint-based causal discovery algorithm; eliminating spurious associations between variables by performing conditional independence tests to infer the structure of a directed acyclic graph describing the causal direction between variables; estimating causal effects using structural equation modeling or a do-calculus framework for directed edges from event chain nodes to load factors based on the obtained causal graph, quantifying the average impact of event chain nodes on load values ​​under controlled confounding factors; extracting all directed paths from the initial event chain trigger point to load variables from the causal graph; calculating the strength of each path, characterized by the product or minimum of the causal effects of each link in the path; and calculating the risk value of each path on the load by combining the path strength with the probability of event chain occurrence, thus achieving a probabilistic assessment of the potential impact on load.

8. The power load forecasting method as described in claim 7, characterized in that, The method further includes: the path strength calculation formula is: for a causal path P:A→B→C, its path strength can be defined as: CE represents the estimated causal effect of each edge; the risk value of each path to the load is assessed by combining path strength and event probability; its calculation formula is: P(A) is the probability of the path starting event A occurring, Strength(P) is the strength of the path, and Impact is the severity coefficient of the load impact.

9. The power load forecasting method as described in claim 7, characterized in that, The method of real-time detection and classification of event chains based on heterogeneous data sources using sequence pattern recognition algorithms also includes: establishing a collaborative prediction mechanism by constructing a distributed information sharing network for microgrid clusters, integrating climate and load data from multiple nodes; designing a prediction coordination algorithm for microgrid clusters, and improving the overall prediction accuracy of the microgrid clusters through dynamic weight allocation; automatically assigning a node with load-sharing capabilities that meets the requirements of a node when an extreme climate event chain is detected by a node in the microgrid cluster based on the event chain detection results; and constructing a multi-dimensional resilience assessment system for microgrid clusters, and monitoring the mutual support capabilities of nodes and the reliability of predictions in real time.

10. The power load forecasting method as described in claim 9, characterized in that, The load capacity specifically includes: a continuous operation event chain detection algorithm for each node in the microgrid group; when a node detects an extreme weather event chain and its risk index exceeds a preset threshold, it broadcasts an event alarm to the group; after receiving the alarm, it verifies the event chain type, impact range, and node load status to determine whether to initiate load diversion; it collects the operating parameters of each node in the group in real time, including the current load factor, maximum carrying capacity, backup power status, line transmission margin, and local event risk level; it uses a multi-index fusion method to calculate the real-time load capacity and diversion priority of each node; it establishes a load diversion optimization model with the goal of minimizing total operating risk, with constraints including node capacity limits, power balance equations, and network security constraints; it solves the optimal load transfer scheme between nodes through a distributed optimization algorithm; it converts the optimal diversion scheme into control commands for execution to ensure a smooth load transition; it monitors the operating status of each node after diversion in real time to verify whether it is in a safe and stable range; the calculation formula for the real-time load capacity of each node is as follows: , This represents the amount of additional load that node i can currently accept. This represents the maximum allowable load for this node. For its current actual load, This is a safety margin calculated based on the local event risk level.

Images