Power supply and demand prediction method adapted to complex market environment

Abstract

The invention relates to the technical field of power supply and demand prediction, and discloses a power supply and demand prediction method suitable for a complex market environment. The method comprises the following steps: collecting multi-dimensional real-time data of a power market, wherein the multi-dimensional real-time data covers a user-side load fluctuation sequence and a market electricity price fluctuation signal; constructing a power supply and demand dynamic prediction model based on the data, and outputting a theoretical supply and demand prediction value; comparing the theoretical predicted value with the actual measurement electric power data, carrying out multi-dimensional difference analysis containing load sequence deviation response delay, and generating a market-level difference coefficient matrix; inputting the matrix into a regional energy network analysis model, generating a supply and demand abnormal propagation path probability distribution diagram and positioning a potential market imbalance region in combination with power grid node transmission capability and economic main body distribution information; and a prediction adjustment strategy is adaptively configured according to the probability distribution diagram, high-frequency data monitoring is started for a high-probability imbalance area, policy disturbance testing is applied to an associated market subject, and stable operation of the power market is supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power supply and demand prediction, in particular to a power supply and demand prediction method suitable for complex market environment. BACKGROUND

[0002] In the current power market environment, the balance of power supply and demand is an important prerequisite for ensuring the stable operation of the power grid and the orderly operation of the market. However, with the continuous development of the power market, its complexity is increasingly prominent, and the traditional power supply and demand prediction method gradually exposes many limitations.

[0003] From the data collection level, most traditional methods only focus on single-dimensional data collection, or focus on user-side load data, or pay attention to power generation-side output data, and lack effective collection and integration of market electricity price fluctuations, which is a key factor affecting supply and demand balance. This single-dimensional data collection method makes the prediction model unable to fully grasp the overall operation situation of the power market, resulting in a large deviation between the prediction result and the actual market situation. The present application collects multi-dimensional real-time data such as user-side load fluctuation sequence, power generation-side output characteristic curve and market electricity price fluctuation signal, which can fully capture key market information.

[0004] In terms of prediction model construction, traditional models are mostly static models, which are difficult to adapt to the dynamic changes of the power market. User-side load presents complex fluctuation sequence affected by factors such as season, weather and industrial production fluctuation, power generation-side output adjusts the output characteristic curve due to factors such as equipment state and fuel supply, and market electricity price fluctuates with real-time changes in supply and demand relationship and policy adjustment. Static models cannot timely capture these changes and are difficult to output theoretical supply and demand prediction values that conform to reality. The present application constructs a power supply and demand dynamic prediction model based on multi-dimensional real-time data, which can timely respond to changes in the user side, the power generation side and the electricity price level, and improve the prediction accuracy.

[0005] In the difference analysis and abnormal positioning link, the traditional method only performs simple numerical comparison, lacks multi-dimensional analysis of load sequence deviation, output curve fitting degree and electricity price signal response delay, and cannot generate a market-level difference coefficient matrix; and it is difficult to combine the power grid node transmission capacity and economic subject distribution information to track the supply and demand abnormal propagation path and locate the potential imbalance area, resulting in lack of pertinence of the control measures. The present application generates a difference coefficient matrix through multi-dimensional difference analysis, locates the imbalance area combined with the regional energy network analysis model, and solves the shortcomings of the traditional method.

[0006] In the regulation strategy, the traditional method adopts unified regulation mode, which cannot formulate differentiated strategy according to the actual situation of different regions. For the region that may be unbalanced, the accurate monitoring mode cannot be enabled in time, and it is also difficult to respond to potential risks through policy disturbance test, so the regulation measures often lag behind market changes and cannot timely resolve the contradiction between supply and demand. The present application adaptively configures the regulation strategy according to the probability distribution diagram, enables high-frequency monitoring for the high-probability unbalanced region, and applies policy disturbance test to the related subject, thereby realizing differentiated regulation. SUMMARY

[0007] The present application aims to provide a power supply and demand prediction method suitable for complex market environment to solve the problems in the above background art.

[0008] To achieve the above-mentioned purpose, the present application provides a power supply and demand prediction method suitable for complex market environment, which comprises:

[0009] Collecting multi-dimensional real-time data of the power market environment, wherein the multi-dimensional real-time data comprises user-side load fluctuation sequence, power generation side output characteristic curve and market electricity price fluctuation signal;

[0010] Constructing a power supply and demand dynamic prediction model based on the multi-dimensional real-time data, and outputting a theoretical supply and demand prediction value;

[0011] Comparing the theoretical supply and demand prediction value with the actual metering power data, performing multi-dimensional difference analysis, wherein the multi-dimensional difference analysis comprises load sequence deviation, output curve coincidence degree and electricity price signal response delay, and generating a difference coefficient matrix at market level;

[0012] Inputting the difference coefficient matrix into a regional energy network analysis model, combining the power grid node transmission capacity and economic subject distribution information, generating a probability distribution diagram of supply and demand abnormal propagation path, and locating the potential market imbalance region;

[0013] Adaptively configuring a prediction regulation strategy according to the probability distribution diagram, including enabling high-frequency market data monitoring mode for the high-probability imbalance region, and applying policy disturbance test to the related market subject.

[0014] Preferably, the multi-dimensional real-time data of the power market environment specifically comprises:

[0015] Performing feature extraction processing on historical market operation data, including extracting long-term trend component and short-term fluctuation component of load fluctuation by using empirical mode decomposition, establishing correlation matrix of electricity price signal and load change through Granger causality analysis, and aligning output curve mode under different market scenarios by using dynamic time warping algorithm;

[0016] The power supply and demand dynamic prediction model construction includes: inputting the processed historical market data into a hybrid prediction network, the hybrid prediction network including an LSTM time series prediction unit based on market cycle characteristics, a fully connected network embedded with volatility attention mechanism, and a prediction weight dynamic adjuster based on real-time data flow;

[0017] The distributed data acquisition unit synchronously captures the ramp rate and peak duration of load fluctuation, the start-stop characteristics and regulation rate of the output curve, and the fluctuation frequency and amplitude gradient of the electricity price signal.

[0018] Preferably, the theoretical supply demand prediction value calculation includes:

[0019] Adaptive filtering based on market environment changes is performed to eliminate measurement noise caused by special events;

[0020] The correlation characteristics of load, output and electricity price are integrated through a multi-source feature fusion algorithm;

[0021] The theoretical prediction value including the normal market fluctuation interval is output, and the theoretical prediction value is dynamically updated according to the market structure change.

[0022] Preferably, the multi-dimensional difference analysis specifically includes:

[0023] Load sequence deviation calculation: sliding comparison of theoretical prediction value and actual measurement value is performed with a variable time window, dynamic time warping algorithm is used to align the non-synchronous data sequences, cumulative deviation amount in each window is calculated, and load deviation vector is generated;

[0024] Output curve fitting degree evaluation: the curve of the theoretical prediction value and the actual output is decomposed, the matching error of the key feature points is calculated, the curve shape similarity index is extracted, and the output fitting degree vector is constructed;

[0025] Electricity price signal response delay detection: analyze the time domain correlation between the theoretical prediction value and the actual electricity price signal, calculate the signal transmission delay and phase shift amount, and generate the electricity price response vector;

[0026] Difference coefficient matrix generation: the load deviation vector, output fitting degree vector and electricity price response vector are fused into a tensor, and the second-order difference coefficient matrix with dimensions of market subjects multiplied by time stamps multiplied by difference types is output through feature weighted normalization processing.

[0027] Preferably, the output curve fitting degree evaluation includes:

[0028] In the curve shape analysis stage, key feature points of the theoretical prediction curve and the actual output curve are extracted, a multi-scale shape matching algorithm is used to match the distribution of the feature points, the shape similarity of a preset key operation interval is calculated, the key operation interval covers a typical market transaction period, after the matching is completed, the shape difference of the theoretical curve and the actual curve in the key operation interval is quantified respectively, the deviation index of each feature point is extracted based on the relative error degree, and the matching results of all feature points are summarized to construct an output coincidence degree vector.

[0029] Preferably, the regional energy network analysis model construction comprises:

[0030] Energy network topology modeling: constructing a regional energy node connection relationship graph according to market subject distribution information, labeling energy transmission parameters between nodes, superimposing reverse power constraint conditions of distributed energy access points in the topology model, and generating a market-oriented topology model containing a transmission capacity matrix and a node sensitivity matrix;

[0031] Abnormal propagation simulation: mapping the difference coefficient matrix to the corresponding nodes of the energy network topology model, performing abnormal propagation deduction based on a graph convolution network, the calculation of the abnormal propagation deduction includes calculating the attenuation factor of abnormal supply and demand according to the node transmission parameters, capturing cross-regional market correlation characteristics through a multi-head attention mechanism, and simulating the diffusion path of abnormal supply and demand in the topology network by using a Monte Carlo method;

[0032] Probability distribution generation: statistics of the abnormal supply and demand appearance frequency of each transmission path in the simulation propagation, calculation of abnormal supply and demand residence probability value combined with transmission capacity parameters, generation of a probability distribution graph covering the whole network, and labeling of a suspicious path set with a probability value exceeding a preset threshold;

[0033] Unbalanced area positioning: performing spatial clustering analysis on the probability distribution graph, identifying abnormal probability aggregation areas, and delineating the physical boundary of market imbalance according to the market subject location and energy network connection relationship.

[0034] Preferably, the adaptive configuration prediction adjustment strategy comprises:

[0035] When the abnormal probability value of a certain area in the probability distribution graph exceeds a preset threshold, a monitoring mode switching instruction is issued to the market unit to which the area belongs, the data acquisition frequency is increased to a specific multiple of the original frequency, the market fluctuation real-time tracking mode is simultaneously enabled, the high-frequency price signal and load mutation are captured, the market event detector is deployed on the edge side, and the extreme price fluctuation segment is recorded;

[0036] The associated market subject with the largest abnormal probability gradient change in the probability distribution graph is subjected to multi-dimensional policy disturbance test.

[0037] Preferably, the multi-dimensional policy disturbance test comprises:

[0038] Inject a multi-frequency test signal through a controllable policy tool, calculate a theoretical market response curve based on a topological model and a transmission capacity matrix, record actual market data of each node after the test signal is injected, obtain a measured market response curve, calculate a difference distance between the theoretical response curve and the measured response curve, compare the difference distance between the theoretical curve and the measured curve, calculate an abnormal node offset percentage, and when the abnormal node offset percentage exceeds a set threshold, mark the policy-sensitive market subject.

[0039] Preferably, the method further comprises:

[0040] Based on the prediction task computing load evaluation, the computing resource is dynamically allocated, the computing load evaluation includes identifying the data magnitude and the computing complexity of the prediction task, when the data volume exceeds the preset threshold, the central prediction system is used as the main computing node, the computing scheduling is carried out through the communication network, the prediction task is dynamically allocated to the auxiliary computing node, and the local computing mode is adaptively switched.

[0041] Preferably, the method further comprises:

[0042] The data security level of the market data is obtained, the data security processing technology is applied based on the data security level, the data use range is limited according to the preset access control strategy, the data security processing technology includes homomorphic encryption and differential privacy, and the access control strategy includes attribute-based access control and role permission management.

[0043] Compared with the prior art, the method has the advantages that:

[0044] From the data acquisition link, the method collects multi-dimensional real-time data of the power market environment, covering user-side load fluctuation sequences, power generation-side output characteristic curves and market price fluctuation signals. This multi-dimensional data acquisition method can fully capture key information in the operation of the power market, breaking the limitations of single-dimensional data acquisition in traditional methods, allowing the prediction model to fully understand the dynamic changes on the user side, the power generation side and the market price level, laying a solid data foundation for subsequent construction of an accurate prediction model, and making the model more in line with the actual operation situation of the power market.

[0045] In the prediction model construction, the power supply and demand dynamic prediction model based on multi-dimensional real-time data is different from the traditional static model and has the ability to adapt to market dynamic changes. Since the input data covers real-time dynamic information of each key link in the market, the model can respond to user-side load fluctuations, power generation-side output adjustments and price changes in a timely manner, thereby outputting more accurate theoretical supply and demand prediction values, greatly improving the reliability of the prediction results and providing more valuable prediction basis for power market participants and regulatory departments.

[0046] In the discrepancy analysis phase, theoretical supply and demand forecasts are compared with actual metered electricity data. A multi-dimensional discrepancy analysis is then performed, incorporating load sequence deviations, output curve alignment, and price signal response delays, ultimately generating a market-level discrepancy coefficient matrix. This multi-dimensional discrepancy analysis approach, compared to traditional simple numerical comparisons, can more comprehensively and deeply reveal the root causes of differences between forecasts and actual values. It clearly presents the operational deviations of the electricity market across multiple key dimensions such as load, output, and electricity prices, providing detailed and systematic discrepancy information to support subsequent anomaly identification and strategy adjustments.

[0047] In terms of anomaly localization, the difference coefficient matrix is ​​input into the regional energy network analysis model, and combined with the transmission capacity of power grid nodes and the distribution information of economic entities, a probability distribution map of the propagation path of supply and demand anomalies is generated, enabling precise localization of potential market imbalance areas. This process fully integrates market difference information, power grid physical characteristics, and the distribution of economic entities, accurately tracking the propagation path of anomalies in the electricity market and identifying areas that may experience imbalances in advance. This changes the situation where traditional methods struggle to accurately locate imbalance areas, allowing regulatory authorities to grasp market risk points in advance and creating favorable conditions for formulating targeted regulatory measures.

[0048] In terms of regulation strategy configuration, an adaptive predictive regulation strategy is configured based on probability distribution maps. High-frequency market data monitoring is activated for high-probability imbalance areas, and policy disturbance tests are applied to related market participants. This adaptive regulation strategy abandons the traditional uniform regulation approach, enabling differentiated measures based on the imbalance risk level of different regions. The high-frequency market data monitoring model can monitor market dynamics in high-risk areas in real time and promptly detect new anomalies; policy disturbance tests can verify the impact of policy adjustments on related market participants in advance, ensuring the effectiveness and rationality of regulatory measures. This more efficiently resolves potential supply and demand contradictions, ensures stable grid operation and orderly operation of the electricity market, and improves the overall operating efficiency and risk resistance of the electricity market. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the working principle of the power supply and demand forecasting method adapted to complex market environments described in this invention.

[0050] Figure 2 A flowchart for building a multi-dimensional data collection and supply-demand dynamic forecasting model;

[0051] Figure 3 Analysis diagram for the implementation of electricity market forecasting methods;

[0052] Figure 4 This is a flowchart for multidimensional difference analysis and the generation of the difference coefficient matrix. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides a method for forecasting power supply and demand in complex market environments. The method collects multi-dimensional real-time data of the power market environment, including user-side load fluctuation sequences, generation-side output characteristic curves, and market electricity price fluctuation signals. Based on this data, a dynamic power supply and demand forecasting model is constructed to output theoretical supply and demand forecasts. Subsequently, the theoretical forecasts are compared with actual metered power data, and multi-dimensional difference analysis is performed to generate a market-level difference coefficient matrix. This matrix is ​​input into a regional energy network analysis model, and combined with the transmission capacity of grid nodes and the distribution information of economic entities, a probability distribution map of the propagation path of supply and demand anomalies is generated, thereby locating potential market imbalance areas. Finally, based on the probability distribution map, an adaptive forecasting and adjustment strategy is configured, including activating a high-frequency market data monitoring mode for high-probability imbalance areas and applying policy disturbance tests to related market entities.

[0055] Example 1: See Figure 2 During the data acquisition phase, feature extraction processing is performed on historical market operation data. The empirical mode decomposition method is used to decompose the load fluctuation sequence into a series of intrinsic mode functions, thereby separating the long-term trend component and the short-term fluctuation component. The long-term trend component reflects the macro-change law of load, while the short-term fluctuation component captures the details of intraday or real-time fluctuations. Granger causality analysis is used to test the causal relationship between electricity price signals and load changes, and a matrix reflecting their dynamic correlation is established. This matrix identifies the guiding relationship between variables under different lag orders. The dynamic time warping algorithm is used to process the output curves under different market scenarios, solving the matching problem caused by inconsistent data acquisition time points or curve shape expansion and contraction, and achieving standardized alignment for diverse operation modes.

[0056] The construction process of the dynamic forecasting model for power supply and demand first involves inputting processed historical market data into a hybrid forecasting network. This network integrates multiple forecasting mechanisms to address market complexity. An LSTM time-series forecasting unit based on market cycle characteristics captures the long-term dependency between load and output, and its gating mechanism effectively learns the fluctuation patterns of the market cycle. A fully connected network with an embedded volatility attention mechanism focuses on abrupt changes and abnormal fluctuations in market data. By assigning differentiated weights to the input data at different time steps, it enhances the model's responsiveness to key signals. A forecasting weight dynamic adjuster based on real-time data streams continuously evaluates the contribution of each forecasting unit and dynamically adjusts the combined weights according to the latest market data to achieve adaptive optimization of the forecasting strategy. Specific indicators of multi-dimensional real-time data are synchronously captured by a distributed data acquisition unit, including the ramp rate and peak duration of load fluctuations to reflect the intensity of demand-side changes, the start-stop characteristics and adjustment rate of the output curve to quantify the response capability of the generation side, and the fluctuation frequency and amplitude gradient of the electricity price signal to characterize market price behavior. These real-time indicators provide continuously updated input information for the model.

[0057] The calculation of theoretical supply and demand forecasts employs adaptive filtering based on changes in the market environment. This processing uses variational mode decomposition combined with wavelet threshold denoising technology to suppress measurement noise caused by special events such as extreme weather or equipment failures, while preserving the true fluctuation characteristics of market data. A multi-source feature fusion algorithm integrates the correlation features of load, output, and electricity price. This algorithm uses a graph neural network-based feature extraction module to mine the spatiotemporal correlations between multi-dimensional data and uses an attention mechanism to fuse feature representations from different sources. The output is a theoretical forecast value that includes the normal market fluctuation range. This value is presented in the form of a probability density function, reflecting the prediction range at different confidence levels. It is dynamically updated through an online learning mechanism as the market structure changes. For example, when new renewable energy is added or market rules are adjusted, the model automatically recalibrates parameters to maintain prediction accuracy.

[0058] See Figure 3This paper demonstrates the complete implementation process of a power supply and demand forecasting method adapted to complex market environments. The left subplot presents the feature extraction and processing results during the data acquisition phase. The load fluctuation sequence is decomposed into multiple intrinsic mode functions (IMFs) and long-term trend components using the empirical mode decomposition (EMD) method. The original load sequence reflects the actual electricity demand changes on the user side, each IMF captures the fluctuation characteristics at different time scales, and the long-term trend component reflects the macroscopic variation law of the load. This multi-scale decomposition method provides rich input features for subsequent forecasting models, effectively separating long-term trends from short-term fluctuation details, and enhancing the model's adaptability to market complexity. The right subplot shows the construction and performance analysis results of the dynamic power supply and demand forecasting model. The figure compares the theoretical forecast load, the actual metered load, and the output results of the LSTM-based time series forecasting unit, while the secondary axis shows the weight changes of the volatility attention mechanism. The LSTM unit effectively learns the fluctuation patterns of market cycles using its gating mechanism, capturing the long-term dependence between load and output; while the attention mechanism enhances the model's response to key signals and abnormal fluctuations by assigning differentiated weights to the input data at different time steps. This hybrid prediction network integrates multiple prediction mechanisms and continuously optimizes the combined strategy through a dynamic adjuster of prediction weights in real-time data streams, achieving adaptive prediction for complex market environments.

[0059] Example 2: See Figure 4 The load sequence deviation is calculated by using a variable time window to compare the theoretical forecast and the actual measurement value. The window size is dynamically adjusted according to the market fluctuation characteristics to adapt to different data frequency requirements. The dynamic time warping algorithm is used to process the asynchronously collected data sequence to solve the alignment problem caused by the difference in sampling time. When calculating the cumulative deviation in each window, both absolute error and relative error are considered to generate a load deviation vector that reflects the deviation change in the time series. In the output curve matching evaluation stage, the theoretical prediction curve and the actual output curve are decomposed into shapes. Wavelet transform and multi-resolution analysis are used to extract key feature points of the curve, including extreme points, inflection points and plateau intervals. The similarity of feature point distribution is compared by a multi-scale shape matching algorithm. This algorithm combines dynamic time warping and pattern recognition technology to calculate the shape similarity of preset key operating intervals. These key operating intervals cover typical market trading cycles such as peak periods, trough periods and transition periods. After matching, the shape differences between the theoretical curve and the actual curve in the key operating intervals are quantified. Hausdorff distance and Fréchet distance are used to measure the curve shape differences. Based on the degree of relative error, the offset index of each feature point, including amplitude deviation and time offset, is calculated. Finally, the matching results of all feature points are summarized to construct the output matching vector.

[0060] The electricity price signal response delay detection and analysis theory predicts the time-domain correlation between electricity price and actual electricity price signals. It uses cross-correlation functions and phase synchronization methods to calculate the transmission delay and phase offset between signals. Instantaneous phase information is extracted through Hilbert transform, and correlation coefficients at different time scales are calculated to identify short-term and long-term response patterns. The generated electricity price response vector includes multi-dimensional indicators such as delay time, offset angle, and correlation coefficient. The difference coefficient matrix generation involves tensor fusion of the previously obtained load deviation vector, output matching vector, and electricity price response vector. High-order tensor decomposition technology is used to integrate difference information from different sources. Feature-weighted normalization is used to assign contribution weights to each vector. Weight allocation is based on entropy weighting and principal component analysis to determine the relative importance of each difference type. The final output dimension is a second-order difference coefficient matrix consisting of market participants multiplied by timestamps and difference types. This matrix comprehensively represents the difference characteristics of different dimensions of the market.

[0061] In the curve morphology analysis stage of the output curve matching evaluation, special attention is paid to the extraction and matching accuracy of key feature points. An improved scale-invariant feature transformation algorithm is used to identify stable feature points in the curve to ensure consistency in identification at different scales. When matching feature point distribution using a multi-scale morphology matching algorithm, rotation invariance and translation invariance processing are introduced. When calculating the morphological similarity of preset key operating intervals, a dynamic programming method is used to find the optimal matching path. The key operating intervals cover typical market trading cycles, including fixed price intervals, bidding intervals, and balanced scheduling intervals. After matching, the morphological differences between the theoretical curve and the actual curve in these intervals are quantified. Based on the degree of relative error, the offset indicators of each feature point are extracted, including position offset, amplitude deviation rate, and morphological distortion. When summarizing the matching results of all feature points, a weighted average method is used to integrate the contributions of different feature points. The final output matching vector can comprehensively reflect the matching quality of curve morphology.

[0062] Taking the actual operation of the East China regional power market from 14:00 to 16:00 on a certain afternoon as an example, a sudden weather change during this period caused a sharp fluctuation in photovoltaic output. The system collected the load fluctuation sequence on the user side during this period, showing that the load in the commercial area decreased by 12% and the load in the residential area increased by 8%. The output characteristic curve on the generation side reflected that the output of the photovoltaic power station decreased by 65% ​​within 30 minutes, and the market electricity price fluctuation signal showed an amplitude oscillation of 3% per minute. The load sequence deviation calculation used a 15-minute variable time window to compare the theoretical prediction value with the actual metering value. The theoretical prediction was generated based on historical data of the same period, and the actual metering value came from smart meters and the data acquisition system. The sequence offset caused by data acquisition delay was aligned by a dynamic time warping algorithm, and the cumulative amount of load deviation within each window was calculated. It was found that the peak load deviation in the commercial area occurred at 14:45, with a deviation of 18.7 MW, and the peak load deviation in the residential area reached 13.2 MW at 15:20. Finally, a load deviation vector containing timestamps and deviation amounts was generated. The output curve matching assessment involves morphological decomposition of the theoretical prediction curve and the actual photovoltaic output curve, extracting key feature points including the starting point of output decline, the lowest valley point, and the recovery inflection point. A multi-scale morphological matching algorithm is used to compare the distribution of feature points and calculate the morphological similarity of the key operating interval from 14:00 to 16:00. It is found that the theoretical curve failed to capture the steep drop in actual output at 14:30, with a shape similarity index of only 0.63. Quantifying the morphological differences between the theoretical and actual curves in the steep drop interval, the feature point offset index shows a time offset of 8 minutes and an amplitude deviation rate of 42%. These are then used to generate an output matching vector.

[0063] The electricity price signal response delay detection and analysis theory predicts the time-domain correlation between electricity price and actual electricity price signals. Using a cross-correlation function to calculate the signal transmission delay, it is found that the electricity price signal response to power output changes has a delay of 12-18 minutes, with phase offset fluctuating between 0.3-0.5 radians. An electricity price response vector incorporating both delay time and phase offset is generated. The difference coefficient matrix generation involves tensor fusion of the load deviation vector, power output matching vector, and electricity price response vector. Weighted normalization is applied to the difference characteristics across four dimensions: commercial area load, residential area load, photovoltaic power output, and electricity price signal. The weight allocation is determined based on the degree of market influence: photovoltaic power output weight is set to 0.4, electricity price signal weight to 0.3, and load weight to 0.15 each. The output is a second-order difference coefficient matrix with dimensions of 4×120×3 (4 market entities × 120 timestamps × 3 difference types).

[0064] In the detailed implementation of the output curve matching evaluation, an improved scale-invariant feature transformation algorithm was used for feature point extraction. Stable feature points in the curve were identified, including points with an output change rate exceeding 10% / minute, sustained low-output plateaus, and rapid recovery phases. A multi-scale morphological matching algorithm incorporated rotation invariance processing to ensure the accuracy of feature point matching across different time scales. Preset key operating intervals covered critical market trading periods, including the pre-decline phase from 14:00 to 14:30, the sharp decline phase from 14:30 to 15:00, and the recovery phase from 15:00 to 16:00. Morphological difference quantification used the Hausdorff distance metric to measure curve shape differences. The calculation showed that the morphological difference was greatest during the sharp decline phase, with a distance value reaching 3.7. Feature point offset indicators included time offset, amplitude deviation rate, and morphological distortion. The final output matching vector contained the matching results of 24 feature points. Data time resolution reached the second level, and data status was synchronized between computing nodes via a high-speed network to ensure the real-time nature and accuracy of the difference analysis. The analysis results were updated every 5 minutes.

[0065] Example 3: The construction of a regional energy network analysis model begins with energy network topology modeling. This process constructs a regional energy node connection diagram based on the distribution information of market entities. In the diagram, nodes represent energy entities such as power plants, load centers, and substations, and edges represent transmission lines or energy transmission channels. Each edge is labeled with energy transmission parameters, including line capacity, impedance characteristics, and power flow constraints. In the topology model, reverse power constraints of distributed energy access points are specially superimposed to reflect the grid connection impact of distributed power sources. Finally, a market-oriented topology model containing a transmission capacity matrix and a node sensitivity matrix is ​​generated. The transmission capacity matrix quantifies the maximum power transmission capacity between nodes, and the node sensitivity matrix characterizes the sensitivity of node power changes to the overall network power flow.

[0066] In the anomaly propagation simulation phase, the difference coefficient matrix is ​​mapped to the corresponding nodes in the energy network topology model. The difference coefficient of each node reflects the degree of its supply and demand anomaly. Anomaly propagation inference is performed based on graph convolutional networks to capture the spatial dependencies between nodes. The calculation of anomaly propagation inference includes calculating the attenuation factor of abnormal supply and demand based on the node transmission parameters. The attenuation factor considers actual engineering parameters such as line loss and transformer turns ratio. A multi-head attention mechanism is used to capture cross-regional market correlation characteristics. This mechanism calculates the correlation weights between different nodes in parallel to identify potential impact paths. The Monte Carlo method is used to simulate the diffusion path of abnormal supply and demand in the topology network. Various possible propagation scenarios are simulated through random sampling to assess system risk. The probability distribution generation process statistically analyzes the frequency of abnormal supply and demand occurrences in each transmission path during simulated propagation. The frequency calculation is based on a large number of repeated experiments in Monte Carlo simulation. Combined with transmission capacity parameters, the abnormal supply and demand residence probability value is calculated. Transmission capacity parameters include line thermal stability limits and voltage stability constraints. A probability distribution map covering the entire network is generated and visualized using a heat map. Suspicious path sets with probability values ​​exceeding a preset threshold are marked. The threshold is dynamically adjusted based on historical fault data and system safety standards. Spatial clustering analysis is performed on the probability distribution map to locate the imbalanced areas. Density clustering algorithm is used to identify abnormal probability clusters. Based on the location of market entities and their connection to the energy network, the physical boundaries of market imbalance are delineated by comprehensively considering electrical distance and power sensitivity. The boundary delineation follows the minimum cut set principle to ensure the mutual independence between regions.

[0067] During the anomaly propagation simulation, the probability of abnormal supply and demand patterns residing along the path is calculated using the following formula:

[0068]

[0069] Where: P ap This represents the probability of abnormal residence, where N is the number of Monte Carlo simulations, and d i This represents the distance at which the anomaly intensity decays in the i-th simulation. This is the network attenuation coefficient. Indicates the duration of the anomaly. This is the system time constant. This formula quantifies the probability and extent of an anomaly's persistence along a specific path.

[0070] During model construction, special attention is paid to the dynamic updating mechanism of the network topology. When new energy facilities are added or the network structure changes, the topology model can automatically update node connection relationships and transmission parameters, ensuring that the analysis results are consistent with the actual power grid state. Anomaly propagation simulation employs a distributed computing architecture to handle large-scale network data, accelerating the Monte Carlo simulation process through parallel computing. Sensitivity analysis is performed after probability distribution generation to assess the impact of changes in key parameters on the probability distribution, improving the reliability of the results. Imbalance area location not only considers the current network state but also incorporates time-dimensional analysis to identify imbalance areas that may migrate over time, providing forward-looking guidance for market monitoring.

[0071] Example 4: The implementation process of the adaptive configuration predictive adjustment strategy is illustrated through a specific example. Suppose that the regional probability distribution map of a certain electricity market shows that the abnormal probability values ​​of three key areas (Area A, Area B, and Area C) are 0.92, 0.87, and 0.43, respectively, with a preset threshold of 0.85. When the abnormal probability values ​​of Area A and Area B exceed this threshold, the system immediately issues a monitoring mode switching instruction to the market units to which these two areas belong, increasing the data acquisition frequency from once per minute to 10 times per minute, and simultaneously enabling the real-time market fluctuation tracking mode. This mode uses a sliding window algorithm to capture high-frequency price signals and load changes, deploys a market event detector based on a rule engine on the edge side, sets a recording mechanism to be triggered when the price volatility exceeds 15%, and saves extreme price fluctuation segments to the local buffer.

[0072] Multi-dimensional policy disturbance tests were conducted on the relevant market entities with the largest abnormal probability gradient changes in the probability distribution map. In the example, area B showed a probability gradient change of 0.12 / hour, which was identified as the area with the highest gradient change. Three major power generation companies in this area were selected as test objects. Multi-frequency test signals were injected through controllable policy tools. The test signals included tiered electricity price subsidy signals, capacity compensation adjustment signals, and environmental subsidy fluctuation signals. The theoretical market response curve was calculated based on the topology model and transmission capacity matrix. Matrix operations considered line transmission limitations and node sensitivity coefficients. After the test signal injection, the actual market data of each node was recorded, and real-time market operation data was obtained at a frequency of once per second, including output adjustment, price change rate, and load response time. The measured market response curve was obtained through data cleaning and time alignment. The difference distance between the theoretical response curve and the measured response curve was calculated. After aligning the time series using a dynamic time warping algorithm, the Euclidean distance was calculated as a difference metric.

[0073] By comparing the difference between the theoretical curve and the measured curve, the percentage offset of abnormal nodes is calculated. A threshold of 20% is set for the percentage offset. When the percentage offset of a coal-fired power plant in Zone B reaches 35%, the system automatically marks the enterprise as a policy-sensitive market entity and uploads this label to the central monitoring system for subsequent policy-making reference. See Table 1 for details of the monitoring parameter adjustments for the three zones.

[0074] Table 1: Adjustment Table of Regional Monitoring Parameters

[0075] During the implementation of the monitoring mode switch, the system adopts a smooth transition mechanism to avoid the impact of sudden changes in data acquisition frequency. First, the acquisition frequency is gradually increased from 1 time / minute to 4 times / minute. After observing the system load, it is further increased to the target frequency. The real-time market fluctuation tracking mode includes two parallel working threads. One thread is dedicated to handling price signal anomaly detection, and the other thread is responsible for load change identification. The two threads exchange abnormal event information through shared memory.

[0076] The market event detector deployed at the edge employs a multi-level caching architecture. The first-level cache stores the raw price data stream, the second-level cache stores preprocessed feature data, and the third-level cache stores identified extreme event fragments. The detector has an adjustable volatility threshold and adjusts triggering conditions in real time based on market volatility. When an extreme event is detected, a timestamp is automatically generated, and high-frequency data from five minutes before and after the event is compressed and stored. Policy disturbance testing is implemented using a gradual strategy. Initially, a small-amplitude test signal is injected to observe the market reaction, then the signal strength is gradually increased. The test signal injection time is chosen during a stable market period to avoid overlap with natural volatility events. The duration of each test signal is controlled between 15 and 30 minutes, and a one-hour data recording window is retained before and after the signal injection for effect comparison. The calculation of the theoretical market response curve uses a physically constrained optimization model, considering the output characteristics of generator sets, line transmission limits, and node balance constraints. The measured market response curve is obtained through a high-precision synchronous phasor measurement device, with time synchronization accuracy reaching the microsecond level, ensuring data comparability.

[0077] The difference distance calculation employs a multi-dimensional measurement method, considering not only curve shape differences but also response time delay, amplitude deviation, and stability indicators. The calculation of the percentage of abnormal node offset integrates both short-term and long-term trend offset components, obtaining the final result through weighted averaging. The labeling of policy-sensitive market entities utilizes a multi-dimensional evaluation system, considering factors such as response consistency, predictability, and stability in addition to the percentage of offset. Labeling results are categorized into three levels: high sensitivity, medium sensitivity, and low sensitivity, with detailed evaluation reports recorded for subsequent analysis. When a monitoring mode switch or policy testing causes system anomalies, the system automatically reverts to its previous safe state. All operations are logged in detail, including operation time, execution parameters, and effect feedback, forming a complete operation traceability chain. The system continuously optimizes monitoring and testing strategies, analyzing historical operation effects through machine learning algorithms to automatically adjust monitoring frequency thresholds, policy testing intensity, and evaluation parameters, improving the accuracy and adaptability of strategy implementation and forming a closed-loop optimization mechanism.

[0078] Example 5: Dynamic Allocation of Computational Resources Based on Forecast Task Computational Load Assessment. A specific example illustrates this process. Assume a provincial power market's day-ahead forecasting task requires processing multi-data sources from smart meters, power generation companies, and the trading platform. The total data volume reaches TB levels and includes various heterogeneous formats. The system first identifies the data volume and computational complexity of the forecasting task. Data volume assessment is based on the number of data points, dimensions, and update frequency. Computational complexity assessment considers the time and space complexity characteristics of the algorithm. When the data volume exceeds a preset single-node processing threshold, the system automatically sets the central forecasting system as the primary computing node. Computational scheduling is performed through the communication network. The primary computing node uses a distributed task scheduling algorithm to decompose the forecasting task into multiple sub-tasks, including data preprocessing, feature extraction, model calculation, and result aggregation. The forecasting task is dynamically allocated to auxiliary computing nodes, including cloud servers and edge computing devices, forming a hierarchical computing architecture. The system adaptively switches between local computing modes. When network bandwidth is insufficient or latency is high, a local degradation computing mode is activated, employing a simplified algorithm to ensure basic forecasting functionality.

[0079] The data security level for acquiring market data is divided into four levels based on data sensitivity and source: Level 1 is public data, Level 2 is internal data, Level 3 is sensitive data, and Level 4 is core data. Data security processing technologies are applied based on these security levels. Homomorphic encryption is applied to Level 3 and above data, allowing predictive calculations to be performed in an encrypted state. Differential privacy technology is applied to data involving personal privacy, adding appropriate noise to prevent individual data identification. Data usage is restricted according to preset access control policies. The access control policy adopts an attribute-based access control model, defining data user identity attributes, environmental attributes, and operational attributes. Combined with role-based access control, different roles are defined, such as system administrators, data analysts, and ordinary users. Each role is assigned differentiated data access permissions and operational permissions, following the principle of least privilege.

[0080] During the implementation of the computational load assessment, the system establishes a real-time monitoring mechanism to continuously collect CPU utilization, memory usage, network bandwidth, and disk I / O performance indicators of each computing node. When an auxiliary computing node is detected to be overloaded, some subtasks are automatically migrated to other idle nodes. The task migration process ensures data consistency and computational continuity. A checkpoint mechanism is used to periodically save the computational state to prevent task interruption. The specific application of data security processing technology reflects multi-level protection. Homomorphic encryption adopts a fully homomorphic encryption scheme, supporting the addition and multiplication homomorphic operations required by the prediction model, ensuring the correctness of the encrypted data calculation results. Differential privacy technology dynamically adjusts the amount of noise added according to the data query frequency and accuracy requirements, balancing data utility and privacy protection. The implementation of access control policies adopts an attribute-based encryption mechanism. The attributes of data users are verified through digital certificates, and a dynamic access token is generated for each data access.

[0081] The communication network employs multiple security mechanisms for computation scheduling. It establishes an end-to-end encrypted channel to transmit computation tasks and data, uses digital signatures to verify the legitimacy of task origins, and uses hash verification to ensure the integrity of task execution results, preventing data tampering during transmission. The scheduling system records complete task allocation logs and computation node status history, supporting post-event auditing and troubleshooting. The adaptive switching to local computation mode includes an intelligent decision-making mechanism. The system assesses network conditions, computational resources, and task urgency in real time. When network latency exceeds a threshold or computational resources are insufficient, it automatically switches to local computation mode. Local computation uses simplified models and compressed data to ensure core prediction functions are completed with limited resources. Once the network or resources are restored, the data is resynchronized to the central system.

[0082] The system dynamically adjusts permissions to control data usage scope. It periodically evaluates user behavior patterns and automatically downgrades permissions when abnormal access is detected. Access to critical data requires multi-factor authentication and approval processes. Operations on high-security data are logged throughout their entire lifecycle, including access time, operation type, and result set size. The system dynamically adjusts computing resources based on the workload of predicted tasks, automatically expanding the number of computing nodes during peak periods and releasing excess resources during off-peak times to save costs. This gradual resource adjustment strategy avoids system fluctuations caused by frequent scaling. The system optimizes resource allocation strategies through machine learning algorithms, analyzes historical task execution data, builds task computation prediction models, pre-allocates necessary resources, reduces task waiting time, improves resource utilization, and forms an intelligent resource management closed loop.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for power supply and demand forecasting that is adaptive to complex market environments, characterized by, The application relates to a method for market imbalance analysis and prediction in a power supply and demand system. The method comprises the following steps: Collecting multidimensional real-time data of a power market environment, including user-side load fluctuation sequences, power generation-side output characteristic curves and market electricity price fluctuation signals; Building a power supply and demand dynamic prediction model based on the multidimensional real-time data, and outputting theoretical supply and demand prediction values; Comparing the theoretical supply and demand prediction values with actual metering power data, performing multidimensional difference analysis including load sequence deviation, output curve fitting degree and electricity price signal response delay, and generating a difference coefficient matrix at a market level; Inputting the difference coefficient matrix into a regional energy network analysis model, combining power grid node transmission capacity and economic subject distribution information, generating a probability distribution diagram of supply and demand abnormality propagation paths, and locating potential market imbalance regions; 2. The method of claim 1, wherein, According to the probability distribution diagram, adaptively configuring a prediction adjustment strategy, including enabling a high-frequency market data monitoring mode for high-probability imbalance regions and applying policy disturbance tests to related market subjects. The collecting of the multidimensional real-time data of the power market environment specifically comprises the following steps: Performing feature extraction processing on historical market operation data, including extracting long-term trend components and short-term fluctuation components of load fluctuation by using empirical mode decomposition, establishing a correlation matrix of electricity price signals and load changes through Granger causality analysis, and aligning output curve modes in different market scenarios by using a dynamic time warping algorithm; The power supply and demand dynamic prediction model building comprises the following steps:

3. The method for power supply and demand forecasting in accordance with complex market environment of claim 2, wherein, Inputting the processed historical market data into a hybrid prediction network, the hybrid prediction network comprising an LSTM time sequence prediction unit based on market cycle characteristics, a fully connected network embedded with fluctuation attention mechanism and a prediction weight dynamic adjuster based on real-time data flow; Synchronously capturing the ramp rate and peak duration of load fluctuation, the start-stop characteristics and adjustment rate of the output curve and the fluctuation frequency and amplitude gradient of the electricity price signal through a distributed data acquisition unit. The theoretical supply and demand prediction value calculation comprises the following steps: Performing adaptive filtering processing based on market environment changes to eliminate measurement noise caused by special events; 4. The method for power supply and demand forecasting in accordance with the complex market environment of claim 1, wherein, Integrating the correlation characteristics of loads, outputs and electricity prices through a multi-source feature fusion algorithm; Outputting theoretical prediction values containing normal market fluctuation intervals, which are dynamically updated with market structure changes. The multidimensional difference analysis specifically comprises the following steps: Load sequence deviation calculation: sliding comparison of the theoretical prediction values and the actual metering values is performed with a variable time window, the dynamic time warping algorithm is used to align the non-synchronously collected data sequences, the cumulative deviation amount in each window is calculated, and a load deviation vector is generated; Output curve fitting degree evaluation: the curve of the theoretical prediction values and the actual output is morphologically decomposed, the matching error of key feature points is calculated, the curve shape similarity index is extracted, and an output fitting degree vector is constructed; Electricity price signal response delay detection: the time domain correlation of the theoretical prediction values and the actual electricity price signals is analyzed, the signal transmission delay and phase shift amount are calculated, and an electricity price response vector is generated. Difference coefficient matrix generation: tensor fusion of the load deviation vector, the output fitness vector and the electricity price response vector, outputting a second-order difference coefficient matrix with dimensions of market subjects multiplied by time stamps multiplied by difference types through characteristic weighted normalization processing.

5. The method for forecasting power supply and demand in accordance with claim 4, wherein, The output curve fitness evaluation includes: In the curve shape analysis stage, key feature points of the theoretical prediction curve and the actual output curve are extracted, a multi-scale shape matching algorithm is used to match the feature point distribution, the shape similarity of a preset key operation interval is calculated, the key operation interval covers a typical market transaction period, after the matching is completed, the shape difference of the theoretical curve and the actual curve in the key operation interval is quantified respectively, the deviation index of each feature point is extracted based on the relative error degree, the matching results of all feature points are summarized, and an output fitness vector is constructed.

6. The method for power supply and demand forecasting in accordance with the complex market environment of claim 1, wherein, The regional energy network analysis model construction includes: Energy network topology modeling: constructing a regional energy node connection relationship diagram according to market subject distribution information, labeling energy transmission parameters between nodes, superimposing reverse power constraint conditions of distributed energy access points in the topology model, and generating a market-oriented topology model including a transmission capacity matrix and a node sensitivity matrix; Abnormal propagation simulation: mapping the difference coefficient matrix to the corresponding nodes of the energy network topology model, performing abnormal propagation deduction based on a graph convolution network, the calculation of the abnormal propagation deduction includes calculating the attenuation factor of abnormal supply and demand according to the node transmission parameters, capturing cross-regional market correlation characteristics through a multi-head attention mechanism, and simulating the diffusion path of abnormal supply and demand in the topology network by using a Monte Carlo method; Probability distribution generation: statistics of the abnormal supply and demand appearance frequency of each transmission path in the simulation propagation, calculation of abnormal supply and demand residence probability values combined with transmission capacity parameters, generation of a probability distribution map covering the entire network, and labeling of a suspicious path set with a probability value exceeding a preset threshold; Unbalanced area positioning: performing spatial clustering analysis on the probability distribution map to identify abnormal probability aggregation areas, and delineating the physical boundary of market imbalance according to the market subject location and energy network connection relationship.

7. The method for power supply and demand forecasting in accordance with the complex market environment of claim 6, wherein, The adaptive configuration prediction adjustment strategy includes: When the abnormal probability value of a certain area in the probability distribution map exceeds a preset threshold, a monitoring mode switching instruction is issued to the market unit to which the area belongs, the data acquisition frequency is increased to a specific multiple of the original frequency, the market fluctuation real-time tracking mode is enabled synchronously, the high-frequency price signal and load mutation are captured, the market event detector is deployed on the edge side, and the extreme price fluctuation segment is recorded; The most relevant market subject with the largest abnormal probability gradient change in the probability distribution map is subjected to multi-dimensional policy disturbance test.

8. The method for power supply and demand forecasting in accordance with the complex market environment of claim 7, wherein, The multi-dimensional policy disturbance test includes: Multi-frequency test signals are injected through controllable policy tools, a theoretical market response curve is calculated based on the topology model and the transmission capacity matrix, actual market data of each node after the test signal injection is recorded to obtain a measured market response curve, the difference distance between the theoretical response curve and the measured response curve is calculated, the difference distance between the theoretical curve and the measured curve is compared, the abnormal node deviation percentage is calculated, and when the abnormal node deviation percentage exceeds a set threshold, it is marked as a policy-sensitive market subject.

9. The method for power supply and demand forecasting in accordance with the complex market environment of claim 1, wherein, Also include: Based on the prediction task computing load evaluation, dynamic allocation of computing resources, the computing load evaluation includes identifying the data magnitude and computational complexity of the prediction task, when the data volume exceeds the preset threshold, the central prediction system is used as the main computing node, the computing scheduling is carried out through the communication network, and the prediction task is dynamically allocated to the auxiliary computing node. Adaptive switching of local computing mode.

10. The method for power supply and demand forecasting in accordance with complex market environment of claim 9, wherein, Also include: Obtain the data security level of market data, apply data security processing technology based on the data security level, limit the data use range according to the preset access control policy, the data security processing technology includes homomorphic encryption and differential privacy, and the access control policy includes attribute-based access control and role permission management.

Images