Data-driven regional power grid load forecasting method, system and medium

Abstract

The application discloses a data-driven regional power grid load prediction method and system and a medium, relates to the technical field of power grids, and comprises the following steps: constructing a prior knowledge base according to a physical power grid topology and a social activity mode; guiding variational mode decomposition of a first load sequence and a prediction task according to the prior knowledge base, and determining a load decomposition item; performing multi-thread prediction on the load decomposition item according to a load prediction plug-in embedded in a power grid management platform, and determining a ternary load prediction sequence; fitting the ternary load prediction sequence, writing a load prediction result into a first register, and executing power distribution guidance. The application solves the technical problem that, in regional power grid load prediction, it is difficult to simultaneously consider long-term trends, periodic laws and the influence of sudden events, resulting in insufficient prediction accuracy and low real-time performance, and achieves the technical effect of improving prediction accuracy and robustness by fusing physical topology and social activity prior knowledge, decomposing load components and implementing multi-thread prediction.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and in particular to a data-driven method, system, and medium for predicting regional power grid load. Background Technology

[0002] With the high proportion of new energy integration, the rapid growth of electric vehicles and controllable loads, and the increasingly frequent interaction between power generation and load in distribution networks, regional power grid loads are exhibiting stronger nonlinear, non-stationary, and stochastic characteristics. Traditional regional power grid load forecasting methods mainly rely on statistical analysis of historical load data and some simple mathematical models, such as time series analysis and regression analysis. These methods can predict the changing trends of power grid loads to a certain extent, but they are gradually revealing many limitations when facing the complex and ever-changing modern power grid environment. On the one hand, the scale of modern regional power grids is constantly expanding, and their structure is becoming increasingly complex, encompassing various types of power generation equipment and power loads. Moreover, the power grids in different regions are interconnected and mutually influential, making the changing patterns of loads more unpredictable. On the other hand, the diversified development of social activity patterns, such as the rise of emerging industries, the holding of large-scale commercial activities, and changes in residents' lifestyles, have all had a significant impact on power grid loads. These factors have a high degree of randomness and uncertainty, and traditional methods cannot fully consider the impact of these complex factors on the load, resulting in a large deviation between the forecast results and the actual load, failing to meet the needs of modern power grid refined management. Summary of the Invention

[0003] This invention provides a data-driven method, system, and medium for regional power grid load forecasting to address the technical problem that it is difficult to simultaneously consider long-term trends, periodic patterns, and the impact of sudden events in regional power grid load forecasting, resulting in insufficient forecast accuracy and low real-time performance. The invention achieves the technical effect of improving forecast accuracy and robustness by integrating prior knowledge of physical topology and social activities, decomposing load components, and implementing multi-threaded forecasting.

[0004] In a first aspect, the present invention provides a data-driven regional power grid load forecasting method, wherein the data-driven regional power grid load forecasting method includes:

[0005] Based on the physical power grid topology and social activity patterns, a prior knowledge base is constructed. Using this prior knowledge base, a variational mode decomposition is performed on the first load sequence and the prediction task to determine load decomposition items. These load decomposition items include at least a long-term trend component, a deterministic periodic component, and an event-driven stochastic component. Using the load prediction plugin embedded in the power grid management platform, multi-threaded prediction is performed on the load decomposition items to determine a ternary load prediction sequence. The load prediction plugin integrates trend prediction threads, periodic prediction threads, and event prediction threads. The ternary load prediction sequence is fitted, and the load prediction results are written to a first register to execute distribution guidance. Event-driven predictions are dynamically fitted with probability.

[0006] Secondly, the present invention also provides a data-driven regional power grid load forecasting system, wherein the data-driven regional power grid load forecasting system includes:

[0007] The system comprises the following modules: a knowledge base construction module (constructing a prior knowledge base based on the physical power grid topology and social activity patterns); a mode decomposition module (guiding variational mode decomposition of the first load sequence and prediction task based on the prior knowledge base to determine load decomposition items, wherein the load decomposition items include at least a long-term trend component, a deterministic periodic component, and an event-driven random component); a multi-threaded prediction module (performing multi-threaded prediction on the load decomposition items based on the load prediction plugin embedded in the power grid management platform to determine a ternary load prediction sequence, wherein the load prediction plugin integrates a trend prediction thread, a periodic prediction thread, and an event prediction thread); and a distribution guidance module (fitting the ternary load prediction sequence, writing the load prediction results into a first register, and executing distribution guidance, wherein event-driven predictions are dynamically fitted with probability.

[0008] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the data-driven regional power grid load forecasting method provided by the present invention.

[0009] This invention discloses a data-driven regional power grid load forecasting method, system, and medium, comprising: constructing a prior knowledge base based on physical power grid topology and social activity patterns; using the prior knowledge base to guide a first load sequence and forecasting task to perform variational mode decomposition to determine load decomposition items, wherein the load decomposition items include at least a long-term trend component, a deterministic periodic component, and an event-driven random component; performing multi-threaded forecasting on the load decomposition items using a load forecasting plugin embedded in a power grid management platform to determine a ternary load forecasting sequence, wherein the load forecasting plugin integrates a trend forecasting thread, a periodic forecasting thread, and an event forecasting thread; fitting the ternary load forecasting sequence, writing the load forecasting results into a first register, and executing distribution guidance, wherein the event-driven forecasting is dynamically fitted with probability. The data-driven regional power grid load forecasting method, system, and medium disclosed in this invention solve the technical problem of insufficient forecasting accuracy and low real-time performance caused by the difficulty in simultaneously considering long-term trends, periodic patterns, and the impact of sudden events in regional power grid load forecasting. It achieves the technical effect of improving forecasting accuracy and robustness by integrating prior knowledge of physical topology and social activities, decomposing load components, and implementing multi-threaded forecasting. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the data-driven regional power grid load forecasting method of the present invention.

[0011] Figure 2 This is a schematic diagram of the structure of the data-driven regional power grid load forecasting system of the present invention.

[0012] Figure labeling: Knowledge base construction module 11, Modal decomposition module 12, Multi-threaded prediction module 13, Power distribution guidance module 14. Detailed Implementation

[0013] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0014] Example 1, as Figure 1 This is a flowchart illustrating the data-driven regional power grid load forecasting method of the present invention, wherein the data-driven regional power grid load forecasting method includes:

[0015] A prior knowledge base is constructed based on the physical power grid topology and social activity patterns.

[0016] Specifically, the process begins with modeling the physical topology of the power grid, including transmission and distribution lines, substations, regional nodes, and their connections. This clarifies the grid's structural hierarchy and power flow path characteristics, and incorporates the capacity constraints, power supply range, and historical load curves of each node into a knowledge base. Next, data related to social activities are collected and organized, including residential and industrial electricity consumption habits, holiday and weekday patterns, meteorological conditions, and the impact of holiday activities on load. Social activity patterns typically exhibit periodicity and regularity; for example, during holidays, the load in commercial and entertainment areas fluctuates significantly due to crowd gatherings. Finally, the acquired grid topology information and social activity pattern information are fused, encoded, and structured for storage, forming a priori knowledge base that incorporates spatial correlation, temporal periodicity, and event-driven patterns. This priori knowledge base can be continuously updated and optimized based on new historical data and actual events, guiding subsequent load decomposition and forecasting processes. This ensures more accurate identification of load change patterns in different regions and social activity scenarios, thereby effectively improving the accuracy of load forecasting.

[0017] Based on the prior knowledge base, the first load sequence and the prediction task are guided to perform variational mode decomposition to determine the load decomposition terms, wherein the load decomposition terms include at least a long-term trend component, a deterministic periodic component, and an event-driven random component.

[0018] Specifically, after constructing the prior knowledge base, the first load sequence is preprocessed by aligning it with the physical power grid topology and social activity patterns in the prior knowledge base. This involves unifying the timestamps to the sampling interval required by the prediction task, supplementing missing data points with the weighted values ​​of adjacent nodes within the partition, and then using a chi-square test to remove gross errors. Subsequently, the number of decomposition modes and the prior center frequency and bandwidth weight of each mode are jointly set using the task-prior approach. The prior center frequency of the trend mode is usually near 0 and a strong bandwidth is set. The prior center frequency of the periodic mode is initialized to be near 2π / 242, 2π / 168 and their harmonics based on the prior period (such as daily / weekly / monthly, zone shifts, price periods, etc.) and is allowed ±5% adaptive drift. The event mode does not preset a narrowband frequency but absorbs anomalies by combining sparsity and high-frequency penalties. Subsequently, augmented Lagrange objectives are constructed using variational mode decomposition (VMD) with prior constraints, and solved using an ADMM iterative approach to update the frequency shift spectrum and time-domain components of each mode. This iteration continues until the relative residual and frequency update amount are both less than the threshold or the maximum number of iterations is reached. Then, modes are classified and merged based on the correlation between the obtained prior center frequency, bandwidth, and prior calendar / meteorological factors. In this process, modes with prior center frequencies close to 0 and coefficients of determination that are weakly correlated with exogenous factors such as temperature, humidity, and holidays are merged into long-term trend components. The coefficients of determination are obtained by performing K-fold robust ridge regression on the mode component and exogenous factors. Modes with prior center frequencies falling within the prior periodic window and showing significant synchronicity with calendar / topology load transfers are summed to form deterministic periodic components. The remaining residual modes that exhibit intermittent high amplitude or are highly correlated with event labels such as holidays, events, maintenance, and power outages are merged into event-driven stochastic components. Finally, by summarizing the long-term trend component, deterministic periodic component, and event-driven random component obtained from the decomposition, a load decomposition term is constructed to ensure that different components in the load sequence are clearly distinguished, providing clear input features for subsequent multi-threaded prediction, which helps to improve the accuracy and interpretability of the prediction results.

[0019] Based on the load forecasting plugin embedded in the power grid management platform, multi-threaded forecasting is performed on the load decomposition items to determine the three-dimensional load forecasting sequence. The load forecasting plugin integrates trend forecasting thread, periodic forecasting thread and event forecasting thread.

[0020] Specifically, a load forecasting plugin is pre-deployed in the power grid management platform. This plugin is embedded into the platform via an interface and possesses parallel computing capabilities. After the load decomposition items are sent to the power grid management platform, the platform uses the long-term trend component, deterministic periodic component, and event-driven random component obtained through variational mode decomposition as inputs, and distributes them to the trend forecasting thread, periodic forecasting thread, and event forecasting thread within the load forecasting plugin. Each thread runs in an independent computing core or virtual thread pool, enabling multi-threaded parallel inference. The trend forecasting thread extrapolates future load changes based on a dynamic factor graph, generating a first-class load forecasting sequence. The periodic forecasting thread calls the periodic memory pool and similarity retrieval mechanism to perform historical pattern matching and time alignment on the decomposed periodic components, and then combines this with the forecasting task to generate a second-class load forecasting sequence. The event forecasting thread calls the constructed event-load mapping relationship and Jacobian determinant to perform probability modeling and conditional distribution adjustment on the input components, generating a third-class load forecasting sequence. Finally, the three threads return the generated three-class sequences to the power grid management platform, which aggregates them into a ternary load forecasting sequence for subsequent power distribution guidance. In summary, this multi-threaded parallel processing mechanism not only shortens the response time of forecast calculations but also ensures that different load components can be modeled and predicted separately under the most suitable model, thereby improving the overall forecast accuracy and robustness and ensuring the real-time performance and reliability of the power grid management platform.

[0021] In some embodiments, before performing multi-threaded forecasting on the load decomposition items, the trend forecasting thread includes:

[0022] The physical power grid topology is partitioned to determine multiple power grid partitions; for the multiple power grid partitions, regional load trend terms are superimposed with exogenous factors as graph nodes; timestamp constraints are introduced, and connection edges are defined by the time-varying correlation between nodes; a dynamic factor graph is determined based on the graph nodes and connection edges, and the dynamic factor graph is initialized based on the first load sequence and the prediction task.

[0023] Specifically, firstly, based on the voltage level, load density, and geographical proximity of the regional physical power grid, spectral clustering is used to partition the physical power grid topology of the target area. In this process, the physical power grid topology is represented as a graph, where nodes represent substations and important load nodes in the power grid, and edges represent transmission and distribution lines. Subsequently, the adjacency matrix and degree matrix of the graph are calculated to obtain a Laplacian matrix. Then, eigenvalue decomposition is performed on the Laplacian matrix, and the eigenvectors corresponding to the k smallest non-zero eigenvalues ​​are taken to form a matrix. The rows of this matrix are normalized, and then the k-means clustering algorithm is used to cluster the normalized row vectors, dividing the nodes into multiple different clusters. Each cluster corresponds to a power grid partition, and each power grid partition contains multiple physical nodes and corresponding user groups. Subsequently, for each power grid zone, the load trend term for that region is extracted, which is the long-term trend component obtained from variational mode decomposition in the previous time period. Exogenous factors such as population size, electricity price level, and policy intensity index are then superimposed onto this trend term, forming a comprehensive feature that can simultaneously reflect physical load inertia and external driving influences. This feature is then used as the node feature in the graph model. Furthermore, to characterize the time-varying coupling relationship between nodes, timestamp constraints are introduced to ensure the consistency of the entire graph within the prediction time window. Within each prediction time window, the Pearson correlation coefficient between the load change rate of nodes in different zones and the influence of exogenous factors is calculated to measure the linear correlation between two nodes. This coefficient is then used as the weight of the connecting edges between nodes, thereby establishing a dynamically changing edge set over time. Based on the determined graph nodes and connecting edges, a dynamic factor graph is constructed. This dynamic factor graph can reflect the time-varying relationship between power grid zones in different time periods. Then, the first load sequence and the prediction task are synchronized to the dynamic factor graph for initialization. That is, the first load sequence is used as the initial state value of the corresponding node, the average Pearson correlation coefficient in the current time window is used as the initial weight of the edge, and the prediction-related parameters in the dynamic factor graph are initialized according to the requirements of the prediction task. For example, if the prediction task is to predict the grid load in the next time period, the prediction time step needs to be set so that it can serve as the core structure support in the subsequent trend prediction and realize the modeling and inference of the load demand in the next time period.

[0024] In some embodiments, performing multi-threaded prediction on the load decomposition items includes:

[0025] The initialized dynamic factor graph is identified, and according to the gating mechanism, the positive and negative diffusion of the influence of exogenous factors is used as the first gating element, and the influence weight of exogenous factors on the load is used as the second gating element to generate a first gating signal. The first gating signal represents the dominant quantity of the load trend, which is either dominated by the load's own inertia or dominated by exogenous factors. Based on the first gating signal, load forecasting based on the dynamic factor graph is performed to determine a type of load forecasting sequence.

[0026] Specifically, after initializing the dynamic factor graph, in order to characterize the transmission effect of exogenous factors in the network, a gating mechanism is used to guide the diffusion effect of exogenous factors. The load trend is encapsulated and calculated through positive and negative influences to construct the first gating signal. This first gating signal is used to represent the dominant factor of the load trend. It is obtained by weighting the first gating element and the first gating signal. The first gating element is divided into positive diffusion and negative diffusion of the influence of exogenous factors. Positive diffusion of the influence of exogenous factors refers to the trend of exogenous factor changes being consistent with the direction of load increase, such as population growth and policies encouraging electricity consumption. Negative diffusion refers to the change of exogenous factors having the opposite effect to load changes, such as rising electricity prices causing some users to actively reduce unnecessary electricity load. The first gating signal includes the influence weight of exogenous factors. This influence weight is automatically adjusted according to different load scenarios. For example, in rigid load scenarios in residential life (such as lighting, cooling, and heating), the weight of the electricity price factor is low because its inhibitory effect is limited. However, in flexible load scenarios (such as electric vehicle charging and interruptible industrial loads), the weight of the electricity price factor is significantly increased, showing a strong inhibitory effect on the total load. When the first gating signal is less than the signal threshold, it indicates that the load itself is dominant, representing that the influence of exogenous factors is insufficient to change the trend. The prediction is mainly based on the extension of historical trend terms, and the ARIMA model is usually used to better capture the time series characteristics and inertial trends of load data. When the first gating signal is greater than the signal threshold, it indicates that exogenous factors are dominant, representing that exogenous factors have a significant driving or inhibitory effect on the load. The prediction is mainly based on the trend after factor adjustment, and the long short-term neural network model is usually used to establish the mapping relationship between exogenous factors and load. Finally, the node states, exogenous factor data, and the first gating signal in the dynamic factor graph are input into the corresponding prediction model to predict the load trend of each node, resulting in a load prediction sequence. This load prediction sequence not only preserves the inertial continuity of the power grid load, but also dynamically reflects the intensity and directionality of the exogenous factors, thus enabling a more realistic description of the changes in the regional power grid load trend in the future time period.

[0027] In some embodiments, multi-threaded prediction is performed on the load decomposition items, wherein the periodic prediction thread includes:

[0028] A periodic memory pool is constructed, wherein the periodic memory pool stores multiple periodic pattern prototypes; a query vector is generated by encoding based on the prediction task and the dynamic factor graph; by calculating the similarity between the query vector and the memory elements in the periodic memory pool, memory element addressing is performed to determine the historical periodic pattern.

[0029] Specifically, firstly, a periodic memory pool is established for each zone on the power grid management platform side, and memory elements are organized according to two-level indexes: period type and seasonal scenario. The period type includes multiple scales such as daily / weekly / monthly / quarterly; the seasonal scenario is represented by meteorological zones and the intensity of social activities. Each memory element stores a periodic pattern prototype, which is represented by a feature vector obtained by standardizing and phase-aligning the historical load sequence within the periodic window, along with exogenous factor statistics, time labels, and confidence levels. Subsequently, the approximate pattern scenario of the first load sequence is extracted from the initialized dynamic factor graph, including the decomposed periodic components and trend residuals of the most recent multiple periodic windows, and the query space is encoded in conjunction with the prediction task. If the prediction task specifies a periodic type such as daily / weekly, a query vector is constructed at the corresponding scale. If a seasonal social state is specified, such as weekdays, summer high temperatures, peak electricity prices, etc., the exogenous factors under that state are introduced as conditional encoding. Next, a joint similarity calculation using cosine similarity and DTW weighted similarity is performed on the query vector and feature vectors of the same period type and seasonal scenario in the memory pool. The feature vector that meets the similarity threshold and has the highest joint similarity is selected. Based on this feature vector, memory cell addressing is performed, and the corresponding memory cell is extracted from the memory pool. The periodic pattern prototype stored in this memory cell is then used as the historical periodic pattern. In summary, this retrieval method can quickly identify historical periodic prototypes similar to the current load scenario and prediction requirements, thus providing a targeted reference pattern for the periodic prediction thread and improving the accuracy and stability of the prediction results.

[0030] In some embodiments, after determining the historical cycle pattern, the process includes:

[0031] Based on the historical periodic pattern, a node time periodicity restoration based on a dynamic factor graph is performed to determine the node periodic information; it is determined whether there is spatial load transfer between nodes, and a correction term is generated; based on the correction term, the node periodic information is compensated to generate a type II load prediction sequence.

[0032] Specifically, after selecting historical cycle patterns, these patterns are mapped to nodes in the dynamic factor graph. Each node then loads a typical load curve from the historical cycle pattern and, combined with the current timestamp and exogenous factor conditions, calculates scaling factors and phases. Based on these scaling factors and phases, the load cycles of each node are aligned and scaled over time to obtain the periodic prediction results for each node, which are recorded as node cycle information. Subsequently, the load flow relationships between nodes in the dynamic factor graph are determined to identify potential spatial load shifts. Spatial load shifts refer to the phenomenon where load from one region is transferred to adjacent regions due to certain factors. For example, during power outages, equipment failures, or grid dispatching, the electricity demand of some regions may be partially or completely transferred to other regions. During the judgment process, the load value at the corresponding time of the previous time period in the first load sequence is calculated by comparing it with the load value in the node periodic information. The absolute value of the calculated difference is then compared with a preset load baseline. If the difference is greater than this load baseline, it indicates that the node has experienced a load anomaly. In this case, it checks whether neighboring nodes exhibit reverse anomalies. If so, the weights between nodes in the dynamic factor graph are enhanced to suppress the absolute value of the correction term. Next, the calculated difference is divided by the sum of 1 and the weights of neighboring nodes to obtain the correction term, which is used to compensate for the periodic pattern distortion caused by spatial transfer of a single node. Finally, the correction term is added to the node periodic information for compensation, forming a second-type load prediction sequence that better reflects the actual spatial coupling relationship. This second-type load prediction sequence not only inherits the stability and interpretability of historical periodic patterns but also corrects the deviation caused by spatial load transfer, thereby improving the accuracy and reliability of periodic prediction and providing a more realistic and effective periodic load reference for power grid dispatch.

[0033] In some embodiments, multi-threaded prediction is performed on the load decomposition items, wherein the event prediction thread includes:

[0034] Retrieve external historical event records and define the event-load mapping relationship, wherein key event parameters affecting the load and load variables are defined; based on the external historical event records, construct the Jacobian determinant by quantifying the local scaling effect of changes in key event parameters on the probability distribution of load variables; construct an event prediction thread based on the event-load mapping relationship and the Jacobian determinant.

[0035] Specifically, the process begins by retrieving historical event records from external sources and establishing a standardized event database. This database includes at least several event types, such as typhoon warnings, cold wave alerts, large-scale events, planned maintenance, and sudden power outages. For each event type, parameters are structured and encoded. Examples include the wind force, duration, and affected area of ​​a typhoon; the minimum temperature and temperature drop of a cold wave; and the number of participants, duration, and location of large-scale events. Subsequently, using an XGBoost or Transformer model, a nonlinear mapping function is trained based on the parameters in the historical event records and the corresponding load data. This trained function is then defined as the event-load mapping relationship. Subsequently, to quantify the local scaling effect of event parameter changes on load distribution, it is assumed that load variables and key event parameters also follow a normal distribution. This maps the event parameter vector to load probability distribution parameters (such as mean and variance), and calculates the partial derivatives of the load distribution function with respect to the event parameters under event disturbances, forming a Jacobian matrix. This Jacobian matrix characterizes the sensitivity of load variables to key event parameters, i.e., the degree of stretching or compression of the load distribution when event intensity changes. Then, based on the event-load mapping relationship and the Jacobian determinant, an event prediction thread is established. During the prediction process, when a potential event is detected in the future time period, this event prediction thread is invoked. The event parameters from the prediction task are input, the mapping relationship is used to locate the direction of load impact, and the Jacobian determinant is used to locally adjust the baseline distribution. This dynamically reflects the unpredictable disturbances of sudden or special events to the regional power grid load, providing a reliable stochastic supplement to the final three types of prediction sequences, ensuring the robustness and continuous operation of the system.

[0036] In some embodiments, the event prediction thread includes:

[0037] For the dynamic factor graph, conditional variables based on the occurrence of external events are determined, wherein the conditional variables are the data distribution under different external events; based on the conditional variables, the event prediction thread is assisted in determining the first probability density based on historical events, and the conditional probability density is determined by adjusting the Jacobian determinant; based on the conditional probability density, the expected load value and confidence interval are determined as three types of load prediction sequences.

[0038] Specifically, firstly, for the dynamic factor graph, external event information is introduced and transformed into conditional variables. These conditional variables characterize the distribution characteristics of load data under different types of external events (such as typhoon warnings, cold wave warnings, large-scale events, etc.). For example, the distribution corresponding to the typhoon conditional variable may show a decrease in industrial load and an increase in residential load, while the cold wave conditional variable corresponds to a significant increase in heating load. These conditional variables are statistically obtained from the historical event database, reflecting the statistical regularity between event occurrence and load change. Subsequently, the event prediction thread uses these conditional variables to perform probabilistic modeling on the historical event data using a Gaussian mixture model (GMM) to obtain the first probability density under different external event scenarios. This first probability density describes the statistical distribution of load levels under a specific event context, including the mean, variance, and distribution tail characteristics. Then, to incorporate real-time event parameters from the prediction task into the modeling, the first probability density is adjusted based on the constructed Jacobian determinant. That is, the local scaling effect of changes in key event parameters on the load distribution is explicitly modeled to obtain the conditional probability density. For example, an increase in typhoon wind force will lead to a greater decrease in industrial load, thus shifting the overall distribution downward and increasing the variance. Finally, based on the corrected conditional probability density, the expected value of the load is calculated as the prediction baseline, and a 95% confidence interval is determined using the quantile method to reflect the probability range of abnormal load fluctuations. By summarizing the calculated expected values ​​and confidence intervals, three types of load prediction sequences are formed. These three types of load prediction sequences are probabilistic prediction objects used to represent the probabilistic causal relationship between possible current emergencies and abnormal load fluctuations. This can help power grid managers better cope with the risks brought by emergencies and load fluctuations, make timely dispatch decisions, and ensure the stable operation of the power grid.

[0039] The three-dimensional load forecast sequence is fitted, the load forecast result is written into the first register, and power distribution guidance is executed. Among them, the event-driven forecast is dynamically fitted with probability.

[0040] Specifically, after obtaining the three-dimensional load forecast sequence, the three types of sequences are aligned and scaled to a unified time base to ensure that forecast results from different sources can be compared and superimposed on the same time axis and under the same dimensions. Then, the first-type load forecast sequence is fused with the second-type load forecast sequence to obtain a baseline load forecast distribution, which characterizes the load probability range without considering the impact of unforeseen events. Afterwards, the three-type load forecast sequences and this baseline load forecast distribution are used as the forecast result and written into the first register within the power grid management platform, ensuring that subsequent dispatching and distribution modules can directly access them. In the distribution guidance stage, the power grid management platform dynamically adjusts the regional power allocation strategy based on the forecast results, including pre-configuring reserve capacity, optimizing power flow distribution, triggering demand response, and contingency plans. This ensures that even in the event of an unforeseen event, the power grid can maintain safety, flexibility, and robustness through rapid switching of forecast corrections and distribution strategies.

[0041] In some embodiments, fitting the ternary load forecast sequence, writing the load forecast result into a first register, and executing distribution guidance include:

[0042] The load forecast sequence of type I and type II are fitted to generate a load forecast distribution; the load forecast distribution and type III load forecast sequence are used as the load forecast result and written into the first register in the power grid management platform. The load forecast distribution is called to provide power distribution guidance, wherein the fitting of the load forecast distribution and type III load forecast sequence is performed according to the event trigger.

[0043] Specifically, firstly, a type I load forecast sequence and a type II load forecast sequence are used as inputs and fitted using a weighted superposition method. The weighting coefficients are dynamically determined in advance based on node importance and the correlation of exogenous factors. After fitting, a continuous base load forecast sequence is obtained. Assuming this base load forecast sequence follows a normal distribution, its probability distribution is fitted as the load forecast distribution. This load forecast distribution includes not only the expected value but also the variance and upper and lower confidence boundaries, used to characterize the load fluctuation range in the future. Subsequently, this load forecast distribution and the type III load forecast sequence (event-driven probability distribution) are stored together as the final load forecast result in the first register of the power grid management platform. The register uses a unified structured data format, storing both the load forecast distribution parameters and the probabilistic descriptions of the three types of forecast sequences, so that they can be retrieved as needed during dispatching or distribution optimization. During distribution guidance, the platform uses the baseline load forecast distribution as the main reference to execute daily power flow optimization, capacity allocation, and dispatching strategies. When an actual event trigger signal is captured, the probability information in the three types of load prediction sequences is automatically called and dynamically weighted and fitted with the load prediction distribution. In this process, the probability distribution convolution is directly performed on the three types of prediction sequences and the load prediction distribution, and the final prediction result is updated in real time. This ensures the stability and accuracy of the prediction under normal scenarios, and can quickly correct the prediction results and adjust the power distribution strategy under emergency scenarios, thereby ensuring the continuity, flexibility and robustness of the regional power grid operation.

[0044] In summary, the data-driven regional power grid load forecasting method provided by this invention has the following technical effects:

[0045] A prior knowledge base is constructed based on the physical power grid topology and social activity patterns. Based on this prior knowledge base, a variational mode decomposition is performed on the first load sequence and the prediction task to determine load decomposition items. These load decomposition items include at least a long-term trend component, a deterministic periodic component, and an event-driven stochastic component. Using a load prediction plugin embedded in the power grid management platform, multi-threaded prediction is performed on the load decomposition items to determine a ternary load prediction sequence. This load prediction plugin integrates trend prediction threads, periodic prediction threads, and event prediction threads. The ternary load prediction sequence is fitted, and the load prediction results are written to a first register for distribution guidance. Event-driven predictions are dynamically fitted with probability, thereby achieving the technical effect of improving prediction accuracy and robustness by integrating prior knowledge of physical topology and social activities, decomposing load components, and implementing multi-threaded prediction.

[0046] Example 2, as Figure 2 This is a schematic diagram of the data-driven regional power grid load forecasting system of the present invention. For example, Figure 1The flowchart of the data-driven regional power grid load forecasting method of the present invention can be illustrated as follows: Figure 2 The structure shown is implemented.

[0047] Based on the same concept as the data-driven regional power grid load forecasting method in the embodiments described above, the present invention also provides a data-driven regional power grid load forecasting system comprising:

[0048] Knowledge base construction module 11: Constructs a prior knowledge base based on the physical power grid topology and social activity patterns; Modal decomposition module 12: Based on the prior knowledge base, guides the first load sequence and prediction task to perform variational mode decomposition to determine load decomposition items, wherein the load decomposition items include at least a long-term trend component, a deterministic periodic component, and an event-driven random component; Multi-threaded prediction module 13: Based on the load prediction plugin embedded in the power grid management platform, performs multi-threaded prediction on the load decomposition items to determine a ternary load prediction sequence, wherein the load prediction plugin integrates a trend prediction thread, a periodic prediction thread, and an event prediction thread; Distribution guidance module 14: Fits the ternary load prediction sequence, writes the load prediction results into a first register, and executes distribution guidance, wherein the event-driven prediction is dynamically fitted with probability.

[0049] In some embodiments, the multi-threaded prediction module 13 includes:

[0050] The physical power grid topology is partitioned to determine multiple power grid partitions; for the multiple power grid partitions, regional load trend terms are superimposed with exogenous factors as graph nodes; timestamp constraints are introduced, and connection edges are defined by the time-varying correlation between nodes; a dynamic factor graph is determined based on the graph nodes and connection edges, and the dynamic factor graph is initialized based on the first load sequence and the prediction task.

[0051] In some embodiments, the multi-threaded prediction module 13 includes:

[0052] The initialized dynamic factor graph is identified, and according to the gating mechanism, the positive and negative diffusion of the influence of exogenous factors is used as the first gating element, and the influence weight of exogenous factors on the load is used as the second gating element to generate a first gating signal. The first gating signal represents the dominant quantity of the load trend, which is either dominated by the load's own inertia or dominated by exogenous factors. Based on the first gating signal, load forecasting based on the dynamic factor graph is performed to determine a type of load forecasting sequence.

[0053] In some embodiments, the multi-threaded prediction module 13 includes:

[0054] A periodic memory pool is constructed, wherein the periodic memory pool stores multiple periodic pattern prototypes; a query vector is generated by encoding based on the prediction task and the dynamic factor graph; by calculating the similarity between the query vector and the memory elements in the periodic memory pool, memory element addressing is performed to determine the historical periodic pattern.

[0055] In some embodiments, the multi-threaded prediction module 13 includes:

[0056] Based on the historical periodic pattern, a node time periodicity restoration based on a dynamic factor graph is performed to determine the node periodic information; it is determined whether there is spatial load transfer between nodes, and a correction term is generated; based on the correction term, the node periodic information is compensated to generate a type II load prediction sequence.

[0057] In some embodiments, the multi-threaded prediction module 13 includes:

[0058] Retrieve external historical event records and define the event-load mapping relationship, wherein key event parameters affecting the load and load variables are defined; based on the external historical event records, construct the Jacobian determinant by quantifying the local scaling effect of changes in key event parameters on the probability distribution of load variables; construct an event prediction thread based on the event-load mapping relationship and the Jacobian determinant.

[0059] In some embodiments, the multi-threaded prediction module 13 includes:

[0060] For the dynamic factor graph, conditional variables based on the occurrence of external events are determined, wherein the conditional variables are the data distribution under different external events; based on the conditional variables, the event prediction thread is assisted in determining the first probability density based on historical events, and the conditional probability density is determined by adjusting the Jacobian determinant; based on the conditional probability density, the expected load value and confidence interval are determined as three types of load prediction sequences.

[0061] In some embodiments, the power distribution guidance module 14 includes:

[0062] The load forecast sequence of type I and type II are fitted to generate a load forecast distribution; the load forecast distribution and type III load forecast sequence are used as the load forecast result and written into the first register in the power grid management platform. The load forecast distribution is called to provide power distribution guidance, wherein the fitting of the load forecast distribution and type III load forecast sequence is performed according to the event trigger.

[0063] In embodiment three, the present invention also provides a computer-readable storage medium that can be used to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the data-driven regional power grid load forecasting method in the embodiments of the present invention, thereby realizing the above-mentioned data-driven regional power grid load forecasting method.

[0064] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A data-driven regional power grid load forecasting method, characterized in that, The method includes: Construct a prior knowledge base based on the physical power grid topology and social activity patterns; Based on the prior knowledge base, the first load sequence and the prediction task are guided to perform variational mode decomposition to determine the load decomposition terms, wherein the load decomposition terms include at least a long-term trend component, a deterministic periodic component and an event-driven random component. Based on the load forecasting plugin embedded in the power grid management platform, multi-threaded forecasting is performed on the load decomposition items to determine the three-dimensional load forecasting sequence. The load forecasting plugin integrates trend forecasting thread, periodic forecasting thread and event forecasting thread. The three-dimensional load forecast sequence is fitted, the load forecast result is written into the first register, and the power distribution guidance is executed. Among them, the event-driven forecast is dynamically fitted with probability. The trend prediction thread, prior to performing multi-threaded prediction on the load decomposition item, includes: The physical power grid topology is partitioned to determine multiple power grid partitions; For the multiple power grid zones, regional load trend terms superimposed with exogenous factors are used as graph nodes; Introduce timestamp constraints to define connection edges based on the time-varying correlation between nodes; Based on the graph nodes and connecting edges, a dynamic factor graph is determined, and the dynamic factor graph is initialized based on the first load sequence and the prediction task. The multi-threaded prediction process for the load decomposition items includes: The dynamic factor graph after initialization is identified. According to the gating mechanism, the positive and negative diffusion of the influence of exogenous factors is used as the first gating element, and the influence weight of exogenous factors on the load is used as the second gating element to generate a first gating signal. The first gating signal represents the dominant quantity of the load trend. The dominant quantity is either dominated by the load's own inertia or dominated by exogenous factors. Based on the first gating signal, load forecasting based on the dynamic factor graph is performed to determine a type of load forecasting sequence.

2. The data-driven regional power grid load forecasting method as described in claim 1, characterized in that, Multi-threaded prediction is performed on the load decomposition items, wherein the periodic prediction thread includes: Construct a periodic memory pool, wherein the periodic memory pool stores multiple periodic pattern prototypes; Based on the prediction task and the dynamic factor graph, a query vector is generated through encoding. By calculating the similarity between the query vector and the memory elements in the periodic memory pool, memory element addressing is performed to determine the historical periodic pattern.

3. The data-driven regional power grid load forecasting method as described in claim 2, characterized in that, After identifying the historical cycle pattern, the following is included: Based on the historical cycle pattern, the node time periodicity is restored based on the dynamic factor graph to determine the node cycle information; Determine if there is spatial load transfer between nodes and generate a correction term; Based on the correction term, the node periodic information is compensated to generate a type II load prediction sequence.

4. The data-driven regional power grid load forecasting method as described in claim 3, characterized in that, Multi-threaded prediction is performed on the load decomposition items, wherein the event prediction thread includes: Retrieve external historical event records and define the event-load mapping relationship, where key event parameters affecting the load and load variables are defined. Based on the external historical event records, a Jacobian determinant is constructed by quantifying the local scaling effect of changes in key event parameters on the probability distribution of load variables. An event prediction thread is constructed based on the event-load mapping relationship and the Jacobian determinant.

5. The data-driven regional power grid load forecasting method as described in claim 4, characterized in that, The event prediction thread includes: For the dynamic factor graph, condition variables based on the occurrence of external events are determined, wherein the condition variables are the data distribution under different external events; Based on the condition variables, the event prediction thread is assisted in determining the first probability density based on historical events, and the condition probability density is determined by adjusting the Jacobian determinant. Based on the conditional probability density, the expected load value and confidence interval are determined as three types of load prediction sequences.

6. The data-driven regional power grid load forecasting method as described in claim 5, characterized in that, Fit the ternary load forecast sequence, write the load forecast results into the first register, and execute distribution guidance, including: The load forecasting sequence of type I and type II are fitted to generate a load forecasting distribution; The load forecast distribution and the three types of load forecast sequences are used as load forecast results and written into the first register in the power grid management platform. The load forecast distribution is called to provide power distribution guidance. The load forecast distribution and the three types of load forecast sequences are fitted according to event triggers.

7. A data-driven regional power grid load forecasting system, characterized in that, The system for implementing the data-driven regional power grid load forecasting method according to any one of claims 1-6, the system comprising: Knowledge base construction module: Construct a prior knowledge base based on the physical power grid topology and social activity patterns; Modal decomposition module: Based on the prior knowledge base, guide the first load sequence and the prediction task to perform variational mode decomposition to determine the load decomposition items, wherein the load decomposition items include at least long-term trend components, deterministic periodic components and event-driven random components. Multi-threaded forecasting module: Based on the load forecasting plugin embedded in the power grid management platform, multi-threaded forecasting is performed on the load decomposition items to determine the three-dimensional load forecasting sequence. The load forecasting plugin integrates trend forecasting thread, periodic forecasting thread and event forecasting thread. Distribution guidance module: Fits the three-dimensional load forecast sequence, writes the load forecast result into the first register, and executes distribution guidance, wherein event-driven forecasts are dynamically fitted with probability.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the data-driven regional power grid load forecasting method as described in any one of claims 1 to 6.

Images