Power load prediction system based on artificial intelligence

Abstract

The invention relates to the technical field of block chains, and discloses a power load prediction system based on artificial intelligence, and the system comprises a data collection module which is used for collecting historical load data, real-time meteorological data, calendar information, and power grid event data; the data storage module is used for performing partition storage on the historical load data, the real-time meteorological data, the calendar information and the power grid event data in the parallel chain to obtain a partition storage result; the feature fusion module is used for performing federal feature extraction on the partition storage result in the parallel chain to obtain federal features and fusing all the federal features; the load prediction module is used for outputting a predicted power load corresponding to the space-time fusion feature through a hybrid expert model; and the load calibration module is used for generating an attribution auxiliary index of the predicted power load and calibrating the predicted power load into the final power load by using the attribution auxiliary index. According to the invention, a new-generation prediction framework considering data security, feature fusion and interpretability can be provided.

Description

Technical Field

[0001] This invention relates to the field of blockchain technology, and in particular to an artificial intelligence-based power load forecasting system. Background Technology

[0002] With the rapid development of smart grids, power load forecasting has become a core technology for ensuring the stable operation of the power grid. Traditional forecasting methods mainly rely on time series models (such as ARIMA) and statistical regression. In recent years, artificial intelligence technologies (especially LSTM, Prophet, and other models) have significantly improved forecast accuracy. To adapt to the need for the fusion of multi-source heterogeneous data, existing research has gradually introduced external factors such as meteorological data, calendar information, and power grid events, attempting to optimize the forecasting effect through feature engineering. At the same time, blockchain technology, due to its advantages of data immutability and distributed storage, has begun to be applied to the security management of power data.

[0003] However, existing technologies still face multiple bottlenecks. First, data sources such as power marketing systems, meteorological centers, and terminal equipment are scattered and independent, forming data silos. Direct centralized storage would pose a risk of privacy leaks, while traditional distributed storage makes it difficult to achieve cross-source feature fusion. Second, due to cross-domain data isolation, most methods can only extract features from a single data source and lack dynamic correlation modeling of cross-dimensional features such as meteorological changes and power grid events. This makes it difficult for models to capture the spatiotemporal coupling patterns of load changes. Third, although mainstream deep learning models can improve accuracy, the output results lack attributability, and maintenance personnel cannot determine whether the prediction deviation is due to meteorological interference or sudden events, making it difficult to implement targeted prediction result calibration.

[0004] In summary, current power load forecasting technology is limited by three core defects: feature fragmentation caused by data silos, lack of cross-domain correlation modeling, and uninterpretable forecast results. These defects not only restrict the improvement of forecast accuracy, but also hinder the reliable application of forecast results in power grid dispatching decisions. There is an urgent need for a new generation of forecasting framework that takes into account data security, feature fusion, and interpretability. Summary of the Invention

[0005] This invention provides an artificial intelligence-based power load forecasting system, the main purpose of which is to provide a new generation forecasting framework that takes into account data security, feature fusion and interpretability.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence-based power load forecasting system, comprising: a data acquisition module, a data storage module, a feature fusion module, a load forecasting module, and a load calibration module; The data acquisition module is used to collect historical load data, real-time meteorological data, calendar information, and power grid event data from the power marketing system, meteorological center, smart meters, and distributed energy terminals, respectively. The data storage module is used to partition and store the historical load data, the real-time meteorological data, the calendar information, and the power grid event data in the parallel chain to obtain the partition storage result; The feature fusion module is used to extract federated features from the partitioned storage results in the parallel chain to obtain federated features, and to fuse the various federated features to obtain spatiotemporal fusion features. The load forecasting module is used to input the spatiotemporal fusion features into a preset hybrid expert model, so as to output the predicted power load corresponding to the spatiotemporal fusion features through the hybrid expert model; The load calibration module is used to generate attribution auxiliary indicators for the predicted power load and to calibrate the predicted power load to the final power load using the attribution auxiliary indicators.

[0007] Optionally, the collection of historical load data, real-time meteorological data, calendar information, and power grid event data from the electricity marketing system, meteorological center, smart meters, and distributed energy terminals includes: Collect hourly electricity consumption by industry segment within historical time periods from the aforementioned electricity marketing system; The hourly electricity consumption of the aforementioned industries is used as historical load data; Temperature data, precipitation data, atmospheric pressure data, and satellite meteorological data are collected from the meteorological center. The temperature data, precipitation data, atmospheric pressure data, and satellite meteorological data are used as real-time meteorological data. Obtain calendar information from the power company's calendar data service; The first power grid event is obtained from the power company's operation and maintenance system; Collect second grid events transmitted back from the smart meters and the distributed energy terminals; The first power grid event and the second power grid event are used as power grid event data.

[0008] Optionally, the step of partitioning and storing the historical load data, the real-time meteorological data, the calendar information, and the power grid event data in the parallel chain to obtain the partitioned storage result includes: Target parallel chains are defined for the historical load data, the real-time meteorological data, the calendar information, and the power grid event data; In each node of the target parallel chain, the corresponding target type data in the historical load data, the real-time meteorological data, the calendar information, and the power grid event data are preprocessed to obtain preprocessed data; In each target parachain, a pre-defined lightweight consensus mechanism is used to verify each pre-processed data and obtain the data verification results. When the data verification result is successful, the preprocessed data fragments are stored in the nodes of the target parallel chain using a partitioned storage method to obtain the original data fragments. The preprocessed data, which is sharded and stored in the nodes of each target parachain, is processed by a hash function to generate a hash digest corresponding to the preprocessed data. The hash digest is synchronized to the main chain through a cross-chain protocol between the parachain and the main chain to construct a hash index matrix corresponding to the hash digest; The original data fragments and the hash index matrix are used as the partition storage results; The target parallel chains include weather chains, load chains, and event chains.

[0009] Optionally, the step of extracting federated features from the partitioned storage results in the parallel chain to obtain federated features includes: Obtain the meteorological chain, load chain, and event chain from the parallel chain; Based on the partitioned storage results, temperature sensitivity coefficients and extreme weather event tags are extracted from the meteorological chain; Based on the partitioned storage results, wavelet transform is used to decompose the historical load data in the load chain to obtain high-frequency and low-frequency components. The instantaneous amplitude and frequency of the historical load data were calculated using the Hilbert-Huang transform. Based on the partitioned storage results, the data in the event chain is encoded as a time window influence factor; The temperature sensitivity coefficient, the extreme weather event label, the high-frequency component, the low-frequency component, the instantaneous amplitude, the frequency, and the time window influence factor are used as federal features.

[0010] Optionally, the fusion of various federated features to obtain spatiotemporal fusion features includes: An attention mechanism is used in the main chain to allocate attention weights for each federated feature; Based on the attention weights, smart contract technology is used in the main chain to aggregate various federated features into spatiotemporal fusion features.

[0011] Optionally, the step of outputting the predicted power load corresponding to the spatiotemporal fusion features through the hybrid expert model includes: Obtain the load dimension features, meteorological dimension features, and event dimension features from the spatiotemporal fusion features; The load dimension features, the meteorological dimension features, and the event dimension features are respectively input into the large model navigation layer, the LSTM expert model, and the Prophet expert model in the hybrid expert model; The long-term load data, meteorological corrections, and event offsets output by the large model navigation layer, the LSTM expert model, and the Prophet expert model are obtained respectively. Based on the meteorological correction and the event offset, the long-term load data is corrected to the predicted power load using the following formula:

[0012] in, Indicates the predicted electricity load. This represents long-term load data. This indicates the amount of weather correction. Indicates the event offset. Represents the meteorological correlation coefficient. This represents the correlation coefficient of the event.

[0013] Optionally, generating the attribution auxiliary index for the predicted power load includes: Calculate the temperature contribution corresponding to the predicted power load; Calculate the event impact value corresponding to the predicted power load; The temperature contribution and the event impact value are used as attribution auxiliary indicators.

[0014] Optionally, calibrating the predicted power load to the final power load using the attribution auxiliary index includes: When the temperature contribution in the attribution auxiliary indicators is greater than a preset temperature threshold and the event impact value in the attribution auxiliary indicators is greater than a preset event threshold, the original parameter data corresponding to the predicted power load is corrected to obtain corrected collected data. The predicted power load is calibrated to the final power load using the corrected collected data.

[0015] Optionally, the step of correcting the original parameter data corresponding to the predicted power load to obtain corrected collected data includes: Initiate a data completion request to the parallel chain where the data supplier corresponding to the original parameter data is located; The data completion request retrieves the completed data from the parachain where the data provider is located. The supplementary data is added to the original parameter data to obtain the corrected acquisition data.

[0016] This invention also proposes an artificial intelligence-based power load forecasting method, characterized in that the method includes: Historical load data, real-time meteorological data, calendar information, and power grid event data are collected from the power marketing system, meteorological center, smart meters, and distributed energy terminals, respectively. The historical load data, real-time meteorological data, calendar information, and power grid event data are partitioned and stored in the parallel chain to obtain the partitioned storage result; In the parallel chain, federated features are extracted from the partitioned storage results to obtain federated features, and the various federated features are fused to obtain spatiotemporal fusion features; The spatiotemporal fusion features are input into a preset hybrid expert model, and the predicted power load corresponding to the spatiotemporal fusion features is output through the hybrid expert model. Generate attribution auxiliary indices for the predicted power load, and use the attribution auxiliary indices to calibrate the predicted power load to the final power load.

[0017] Compared to the background technology, this invention, by introducing external factors such as meteorological data, calendar information, and power grid events, can more comprehensively characterize the influencing factors of power load. Compared to traditional prediction methods that rely solely on historical load data, it can significantly improve prediction accuracy. Data collection from the power marketing system, meteorological center, smart meters, and distributed energy terminals is the premise and foundation for breaking down data silos. This invention, through a parallel chain partitioned storage method, partitions data according to the source department. Combined with a lightweight consensus mechanism, it achieves data sharing and solves data silos while ensuring data security and privacy, avoiding the privacy leakage risks that may be caused by direct centralized storage. Compared to traditional distributed storage, it better balances data sharing and privacy protection. The parallel chain architecture and lightweight consensus mechanism can improve data processing efficiency and provide efficient data support for subsequent feature extraction and fusion. This invention, through federated feature extraction, can provide support for subsequent analysis of meteorological changes, power grid events, etc. This invention utilizes cross-dimensional feature dynamic correlation modeling to prepare for the fusion of multi-source heterogeneous data, thereby helping to capture the spatiotemporal coupling patterns of power load changes. It overcomes the limitations of most existing methods that can only extract features from a single data source. This invention combines the advantages of a large model navigation layer for mining long-term load patterns with a group of expert models for handling short-term fluctuations and special date effects. It can capture both long-term trends in power load and handle short-term fluctuations and special events. Compared to single models (such as LSTM or Prophet alone), it has higher prediction accuracy and better adaptability, and can output more accurate predicted power load. This invention also generates attribution auxiliary indicators such as temperature contribution and event impact values, enabling maintenance personnel to clearly understand the sources of prediction deviations. This solves the problem of lack of attributability in the output results of mainstream deep learning models, allowing for targeted calibration of prediction results and improving the credibility and usability of prediction results in power grid dispatching decisions. Therefore, this invention can provide a new generation of prediction framework that balances data security, feature fusion, and interpretability. Attached Figure Description

[0018] Figure 1 A functional block diagram of an artificial intelligence-based power load forecasting system provided in an embodiment of the present invention; Figure 2 A schematic diagram of load forecasting for an artificial intelligence-based power load forecasting system provided in an embodiment of the present invention; Figure 3 A schematic diagram of load calibration for an artificial intelligence-based power load forecasting system provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating an artificial intelligence-based power load forecasting method according to an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0020] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0021] In practice, the server-side equipment deployed in an AI-based power load forecasting system may consist of one or more devices. This AI-based power load forecasting system can be implemented as: a business instance, a virtual machine, or hardware devices. For example, the AI-based power load forecasting system can be implemented as a business instance deployed on one or more devices in a cloud node. Simply put, this AI-based power load forecasting system can be understood as software deployed on a cloud node, used to provide AI-based power load forecasting services to various user terminals. Alternatively, the AI-based power load forecasting system can also be implemented as a virtual machine deployed on one or more devices in a cloud node. This virtual machine contains application software for managing various user terminals. Or, the AI-based power load forecasting system can also be implemented as a server composed of numerous identical or different types of hardware devices, with one or more hardware devices configured to provide AI-based power load forecasting services to various user terminals.

[0022] In terms of implementation, the AI-based power load forecasting system and the user terminal are mutually compatible. That is, if the AI-based power load forecasting system is implemented as an application installed on a cloud service platform, then the user terminal is a client that establishes a communication connection with the application; or if the AI-based power load forecasting system is implemented as a website, then the user terminal is implemented as a webpage; or if the AI-based power load forecasting system is implemented as a cloud service platform, then the user terminal is implemented as a mini-program in an instant messaging application.

[0023] Reference Figure 1 The diagram shown is a functional block diagram of an artificial intelligence-based power load forecasting system provided in an embodiment of the present invention.

[0024] The artificial intelligence-based power load forecasting system 100 described in this invention can be installed on a cloud server. In terms of implementation, it can be used as one or more service devices, or as an application installed in the cloud (e.g., an AI-based power load forecasting server, server cluster, etc.), or it can be developed as a website. Depending on the functions implemented, the AI-based power load forecasting system 100 includes a data acquisition module 101, a data storage module 102, a feature fusion module 103, a load forecasting module 104, and a load calibration module 105.

[0025] In this embodiment of the invention, in the tracking of AI-based power load forecasting, each of the above modules can be implemented independently and called upon other modules. This "calling" can be understood as one module connecting to multiple modules of another type and providing corresponding services to those connected modules. In the AI-based power load forecasting system provided by this embodiment, the applicability of the AI-based power load forecasting architecture can be adjusted by adding modules and directly calling them without modifying the program code, achieving cluster-based horizontal expansion to quickly and flexibly expand the AI-based power load forecasting system. In practical applications, the above modules can be set in the same device or different devices, or they can be set in virtual devices, such as service instances in a cloud server.

[0026] The following describes the components and specific workflow of an artificial intelligence-based power load forecasting system, using specific embodiments as examples.

[0027] The data acquisition module 101 is used to collect historical load data, real-time meteorological data, calendar information, and power grid event data from the power marketing system, meteorological center, smart meters, and distributed energy terminals, respectively.

[0028] This invention, by introducing external factors such as meteorological data, calendar information, and power grid events, can more comprehensively characterize the factors influencing power load. Compared with traditional forecasting methods that rely solely on historical load data, it can significantly improve forecast accuracy. Data is collected from the power marketing system, meteorological center, smart meters, and distributed energy terminals, which is the premise and foundation for breaking down data silos.

[0029] In one embodiment of the present invention, the step of collecting historical load data, real-time meteorological data, calendar information, and power grid event data from the power marketing system, meteorological center, smart meters, and distributed energy terminals respectively includes: collecting hourly electricity consumption by industry within a historical period from the power marketing system; using the hourly electricity consumption by industry as historical load data; collecting temperature data, precipitation data, atmospheric pressure data, and satellite meteorological data from the meteorological center; using the temperature data, precipitation data, atmospheric pressure data, and satellite meteorological data as real-time meteorological data; obtaining calendar information from the power company's calendar data service; obtaining a first power grid event from the power company's operation and maintenance system; collecting a second power grid event transmitted back by the smart meters and the distributed energy terminals; and using the first power grid event and the second power grid event as power grid event data.

[0030] The electricity marketing system refers to an information system used to manage electricity marketing operations, covering electricity sales, electricity bill settlement, customer service, and other related businesses. For example, a power company's marketing system records each user's monthly electricity consumption and payment status. The industry-specific hourly electricity consumption refers to the hourly electricity consumption data of various industries over different time periods in the past. For example, the hourly electricity consumption data of industries such as industry, commerce, and residential use over the past year. The meteorological center refers to an institution responsible for meteorological monitoring, forecasting, and services, providing various types of meteorological data. For example, the National Meteorological Center releases nationwide weather forecasts and real-time meteorological data. The temperature data refers to... Data reflecting changes in ambient temperature is typically expressed in degrees Celsius or Fahrenheit. For example, the hourly temperature recorded by a weather station in a certain area. Precipitation data reflects rainfall or snowfall, usually expressed in millimeters. For example, 10 millimeters of precipitation in a certain area on a certain day indicates light rain. Atmospheric pressure data reflects the pressure exerted by the atmosphere on the ground or a certain plane, usually expressed in hectopascals (hPa). For example, an atmospheric pressure of 1013.25 hPa in a certain area at a certain time indicates that the air pressure in that area is at a standard level. Satellite meteorological data refers to remote sensing data of the atmosphere, clouds, and surface acquired by meteorological satellites, which can be used for weather forecasting. Monitoring and forecasting, for example, satellite cloud images can show the distribution and movement of clouds, helping to predict weather changes. The calendar data service provides date, time, and related calendar information, typically including holiday and weekend markers. Calendar information refers to various date-related information, such as whether it is a holiday or a workday. The power company's operation and maintenance system refers to the system used by the power company to manage the operation and maintenance of power equipment, recording equipment operating status, maintenance plans, etc. For example, a power company's operation and maintenance system records the regular maintenance plan for equipment in a substation. The first grid event refers to a grid event obtained from the power company's operation and maintenance system, such as equipment... The terms "fault," "maintenance plan," etc., refer to the following: The smart meter is a meter with automatic metering and data transmission functions, capable of recording and uploading users' electricity consumption data in real time. For example, a user's household meter records electricity consumption every 15 minutes and uploads it to the power company in real time. The distributed energy terminal refers to terminal equipment in a distributed energy system, such as solar panels or small wind turbines, used to collect and upload operational data. For example, a company's rooftop solar power generation system's terminal equipment records power generation and equipment operating status. The second grid event refers to grid event data transmitted back from the smart meter and distributed energy terminal, such as equipment failure or abnormal power generation.

[0031] The data storage module 102 is used to partition and store the historical load data, the real-time meteorological data, the calendar information, and the power grid event data in the parallel chain to obtain the partition storage result.

[0032] This invention utilizes a parachain-based partitioned storage method, enabling data to be partitioned according to its source department. Combined with a lightweight consensus mechanism, it achieves data sharing and resolves data silos while ensuring data security and privacy. This avoids the privacy leakage risks that may arise from direct centralized storage. Compared with traditional distributed storage, it better balances data sharing and privacy protection. The parachain architecture and lightweight consensus mechanism can improve data processing efficiency and provide efficient data support for subsequent feature extraction and fusion.

[0033] In one embodiment of the present invention, the step of partitioning and storing the historical load data, real-time meteorological data, calendar information, and power grid event data in a parallel chain to obtain a partitioned storage result includes: dividing the historical load data, real-time meteorological data, calendar information, and power grid event data into target parallel chains; performing data preprocessing on the target type data corresponding to the historical load data, real-time meteorological data, calendar information, and power grid event data in the nodes of each target parallel chain to obtain preprocessed data; and using a preset lightweight consensus mechanism to verify each preprocessed data in each target parallel chain. The data verification results are obtained. When the data verification result indicates successful data verification, each preprocessed data shard is stored in the nodes of each target parachain using a partitioned storage method to obtain the original data shards. The preprocessed data shards stored in the nodes of each target parachain are processed by a hash function to generate a hash digest corresponding to the preprocessed data. The hash digest is synchronized to the main chain through a cross-chain protocol between the parachain and the main chain to construct a hash index matrix corresponding to the hash digest. The original data shards and the hash index matrix are used as the partitioned storage result. The target parachains include a weather chain, a load chain, and an event chain.

[0034] The lightweight consensus mechanism refers to an efficient consensus algorithm that consumes less computational resources compared to traditional Proof of Work mechanisms. It is suitable for networks with relatively high trust between nodes. For example, Practical Byzantine Fault Tolerance (PBFT) is a lightweight consensus mechanism commonly used in blockchain networks, enabling rapid consensus and verification of transaction or data validity. The partitioned storage method refers to distributing different types of data (including historical load data, real-time weather data, calendar information, and power grid event data) across different parallel chains (including weather chains, load chains, and event chains) within a parachain architecture. Within each parachain, data is further distributed across multiple nodes. Original data sharding refers to storing the same type of data on different nodes within the same parachain; for example, data containing user electricity consumption records might be sharded. Storage: Each shard contains records for a specific region or time period. The hash index matrix refers to a data structure that stores hash values ​​for quick data retrieval and location. For example, the hash index matrix on the main chain stores the hash digests of data from various parachains. This matrix allows for quick retrieval of the corresponding parachain data. The main chain refers to the core chain in the blockchain network, which serves as the infrastructure of the entire network and is responsible for coordinating and managing the operation of parachains. The cross-chain protocol refers to a set of rules and standards that allow communication and data exchange between different blockchains. It should be noted that the calendar information and the power grid event data are jointly stored in the event chain. When the data verification result is a data verification failure, the data is re-collected and stored in the parachain.

[0035] Optionally, the process of preprocessing the historical load data, real-time meteorological data, calendar information, and power grid event data corresponding to the target type data in the nodes of each target parachain to obtain preprocessed data refers to operations such as cleaning, format conversion, and handling missing values ​​of the original data to make the data meet the requirements of analysis and modeling. Further, the process of using a preset lightweight consensus mechanism to verify the preprocessed data in each target parachain refers to checking the accuracy, completeness, and consistency of the data to ensure that the data conforms to the expected format and business logic. For example, in the parachain, the lightweight consensus mechanism verifies the data to ensure that the data version of each node is consistent and has not been tampered with. Further, the process of processing the target type data stored in the nodes of each target parachain with a hash function refers to converting data of arbitrary length into a fixed-length hash value through a hash algorithm to uniquely identify the data and verify the integrity of the data. For example, using the SHA-256 hash function to process a piece of text to obtain a 256-bit hash value.

[0036] The feature fusion module 103 is used to extract federated features from the partitioned storage results in the parallel chain, obtain federated features, and fuse the various federated features to obtain spatiotemporal fusion features.

[0037] This invention, through federated feature extraction, can prepare for subsequent dynamic correlation modeling of cross-dimensional features such as meteorological changes and power grid events, and realize the fusion of multi-source heterogeneous data. This helps to capture the spatiotemporal coupling pattern of power load changes and overcomes the limitation of most existing methods that can only extract features from a single data source.

[0038] In one embodiment of the present invention, the step of extracting federated features from the partitioned storage results in the parallel chain to obtain federated features includes: acquiring the meteorological chain, load chain, and event chain in the parallel chain; extracting temperature sensitivity coefficients and extreme weather event labels from the meteorological chain based on the partitioned storage results; performing wavelet decomposition on the historical load data in the load chain using wavelet transform based on the partitioned storage results to obtain high-frequency and low-frequency components; calculating the instantaneous amplitude and frequency of the historical load data using Hilbert-Huang transform; encoding the data in the event chain as time window influence factors based on the partitioned storage results; and using the temperature sensitivity coefficient, the extreme weather event label, the high-frequency component, the low-frequency component, the instantaneous amplitude, the frequency, and the time window influence factor as federated features.

[0039] The high-frequency component refers to the frequency reflecting short-term random fluctuations of the load, such as a frequency >0.5Hz. The low-frequency component refers to the frequency reflecting long-term trends of the load, such as a frequency <0.1Hz. The Hilbert-Huang Transform is a method for analyzing non-stationary and nonlinear signals. It consists of two parts: Empirical Mode Decomposition (EMD) and Hilbert Transform (HHT). The main purpose of HHT is to decompose complex signals into a set of simple, physically meaningful oscillation modes, thereby extracting the time-frequency characteristics of the signal. The instantaneous amplitude refers to the amplitude intensity of the load fluctuation at a certain moment. The frequency refers to the oscillation rate of the load fluctuation at a certain moment. The time window influence factor refers to the time-sensitive weight obtained by encoding the event data in the event chain. Its value rules are as follows: a time window influence factor of 1.0 indicates that the event is within the effective period (such as the period from New Year's Eve to the seventh day of the Lunar New Year during the Spring Festival holiday, with the greatest impact); 0.2 indicates that the event is in the buffer period (the transition period before and after the effective period); and 0 indicates that the event is not activated (not in the effective period or buffer period, with the least impact).

[0040] Optionally, the process of extracting temperature sensitivity coefficients and extreme weather event labels from the meteorological chain based on the partitioned storage results refers to the process of finding the chain node where the original data shard is located through a hash index and processing the original data shard. The temperature sensitivity coefficient is the ratio between the load change and the temperature change, representing the intensity of the impact of temperature change on the load. For example, the temperature sensitivity coefficient of residential load is 5.2 MW / ℃, indicating that for every 1℃ increase in temperature, the load increases by 5.2 MW. The extreme weather event label refers to the discrete code that marks the extreme weather type in the meteorological chain, represented by a one-hot vector. For example, the code for a typhoon is [1,0,0], and the code for a cold wave is [0,1,0]. Further, the process of performing wavelet decomposition on the historical load data in the load chain using wavelet transform based on the partitioned storage results to obtain high-frequency and low-frequency components, and the process of encoding the data in the event chain as time window influence factors based on the partitioned storage results, are all processes of finding the chain node where the original data shard is located through a hash index and processing the original data shard.

[0041] In one embodiment of the present invention, the process of fusing various federated features to obtain spatiotemporal fusion features includes: allocating attention weights for each federated feature in the main chain using an attention mechanism; and aggregating the various federated features into spatiotemporal fusion features in the main chain using smart contract technology based on the attention weights.

[0042] Optionally, the allocation of attention weights for each federated feature using an attention mechanism requires the use of a formula. , Represents the i-th federal feature. Indicates the target forecast period. This represents the activation function. This indicates the importance between input features and target output. Furthermore, the process of aggregating various federated features into spatiotemporal fusion features in the main chain using smart contract technology based on the attention weights refers to using a programming language that supports smart contracts (such as Solidity) to write smart contract code, deploying the written smart contract to the main chain. This process requires utilizing the development tools and environment provided by the blockchain platform, paying a certain deployment fee to complete the storage and deployment of the contract on the main chain, defining the input parameters (various federated features and their corresponding attention weights) and output parameters (spatiotemporal fusion features), and implementing feature aggregation in the code. For example, according to the formula fusion feature = w1 * feature 1 + w2 * feature 2 ..., i.e., performing a weighted summation operation.

[0043] The load prediction module 104 is used to input the spatiotemporal fusion features into a preset hybrid expert model, so as to output the predicted power load corresponding to the spatiotemporal fusion features through the hybrid expert model.

[0044] This invention combines the advantages of a large model navigation layer for mining long-term load patterns and a group of expert models for handling short-term fluctuations and special date effects by using a hybrid expert model. It can capture the long-term trend of power load and handle short-term fluctuations and special events. Compared with a single model (such as LSTM or Prophet alone), it has higher prediction accuracy and better adaptability, and can output predicted power load more accurately.

[0045] In one embodiment of the present invention, the step of outputting the predicted power load corresponding to the spatiotemporal fusion features through the hybrid expert model includes: acquiring load dimension features, meteorological dimension features, and event dimension features from the spatiotemporal fusion features; inputting the load dimension features, meteorological dimension features, and event dimension features into the large model navigation layer, LSTM expert model, and Prophet expert model in the hybrid expert model, respectively; acquiring the long-term load data, meteorological correction, and event offset output by the large model navigation layer, LSTM expert model, and Prophet expert model, respectively; and correcting the long-term load data to the predicted power load using the following formula based on the meteorological correction and the event offset:

[0046] in, Indicates the predicted electricity load. This represents long-term load data. This indicates the amount of weather correction. Indicates the event offset. Represents the meteorological correlation coefficient. This represents the correlation coefficient of the event.

[0047] The large model navigation layer, LSTM expert model, and Prophet expert model are respectively composed of Transformer, Long Short-Term Memory Neural Network, and Prophet. It should be noted that the parameters... , These two parameters are obtained from meteorological data and event data, for example... , Meteorological features of dimensions 13-25 pass To obtain, that is for Output The probability value corresponding to the most significant dimension. These are event dimension features of dimensions 26-38.

[0048] For example, regarding load dimension features, meteorological dimension features, and event dimension features, for instance, dimensions 1-12 of the spatiotemporal fusion features are load dimension features, dimensions 13-25 are meteorological dimension features, and dimensions 26-38 are event dimension features.

[0049] The load calibration module 105 is used to generate attribution auxiliary indicators for the predicted power load and to calibrate the predicted power load to the final power load using the attribution auxiliary indicators.

[0050] This invention, by generating attribution auxiliary indicators such as temperature contribution and event impact value, enables operation and maintenance personnel to clearly understand the source of prediction deviations, solving the problem of lack of attributability in the output results of mainstream deep learning models. This allows for targeted calibration of prediction results, improving the credibility and usability of prediction results in power grid dispatching decisions.

[0051] In one embodiment of the present invention, generating the attribution auxiliary index for the predicted power load includes: calculating the temperature contribution corresponding to the predicted power load using the following formula:

[0052] in, Indicates the contribution of temperature. Indicates the predicted electricity load. This indicates the amount of weather correction. Indicates the meteorological correlation coefficient; The event impact value corresponding to the predicted power load is calculated using the following formula:

[0053] in, Indicates the impact value of the event. Represents event-dimensional features. Represents the correlation coefficient of an event; The temperature contribution and the event impact value are used as attribution auxiliary indicators.

[0054] See Figure 2 The diagram shown is a load forecasting schematic of an artificial intelligence-based power load forecasting system according to an embodiment of the present invention. Figure 2 In the calculation formula, the prediction error is (predicted power load value - actual power load value) / actual power load value * 100%. For example, the prediction error calculation formula in the first row is (258.7-256.3) / 256.3*100%=0.936%.

[0055] In one embodiment of the present invention, calibrating the predicted power load to the final power load using the attribution auxiliary index includes: when the temperature contribution of the attribution auxiliary index is greater than a preset temperature threshold and the event impact value of the attribution auxiliary index is greater than a preset event threshold, performing data correction on the original parameter data corresponding to the predicted power load to obtain corrected acquisition data; and calibrating the predicted power load to the final power load using the corrected acquisition data.

[0056] In another embodiment of the present invention, the step of correcting the original parameter data corresponding to the predicted power load to obtain corrected collected data includes: initiating a data completion request to the parallel chain where the data supplier corresponding to the original parameter data is located; retrieving the completion data from the parallel chain where the data supplier is located through the data completion request; and adding the completion data to the original parameter data to obtain corrected collected data.

[0057] See Figure 3 The diagram shown is a load calibration schematic of an artificial intelligence-based power load forecasting system according to an embodiment of the present invention.

[0058] Optionally, retrieving the supplementary data from the parallel chain where the data supplier is located via the data completion request refers to retrieving the data at the time when the original parameter data corresponding to the predicted power load based on the aforementioned spatiotemporal fusion feature is located. In other words, the original parameter data is retrieved again. The original parameter data here refers to the aforementioned historical load data, real-time meteorological data, calendar information, power grid event data, etc. Adding the supplementary data to the original parameter data to obtain the corrected collected data means comparing the newly retrieved original parameter data with the original parameter data at the beginning, and adding new data where it is missing.

[0059] Compared to the background technology, this invention, by introducing external factors such as meteorological data, calendar information, and power grid events, can more comprehensively characterize the influencing factors of power load. Compared to traditional prediction methods that rely solely on historical load data, it can significantly improve prediction accuracy. Data collection from the power marketing system, meteorological center, smart meters, and distributed energy terminals is the premise and foundation for breaking down data silos. This invention, through a parallel chain partitioned storage method, partitions data according to the source department. Combined with a lightweight consensus mechanism, it achieves data sharing and solves data silos while ensuring data security and privacy, avoiding the privacy leakage risks that may be caused by direct centralized storage. Compared to traditional distributed storage, it better balances data sharing and privacy protection. The parallel chain architecture and lightweight consensus mechanism can improve data processing efficiency and provide efficient data support for subsequent feature extraction and fusion. This invention, through federated feature extraction, can provide support for subsequent analysis of meteorological changes, power grid events, etc. This invention utilizes cross-dimensional feature dynamic correlation modeling to prepare for the fusion of multi-source heterogeneous data, thereby helping to capture the spatiotemporal coupling patterns of power load changes. It overcomes the limitations of most existing methods that can only extract features from a single data source. This invention combines the advantages of a large model navigation layer for mining long-term load patterns with a group of expert models for handling short-term fluctuations and special date effects. It can capture both long-term trends in power load and handle short-term fluctuations and special events. Compared to single models (such as LSTM or Prophet alone), it has higher prediction accuracy and better adaptability, and can output more accurate predicted power load. This invention also generates attribution auxiliary indicators such as temperature contribution and event impact values, enabling maintenance personnel to clearly understand the sources of prediction deviations. This solves the problem of lack of attributability in the output results of mainstream deep learning models, allowing for targeted calibration of prediction results and improving the credibility and usability of prediction results in power grid dispatching decisions. Therefore, this invention can provide a new generation of prediction framework that balances data security, feature fusion, and interpretability.

[0060] like Figure 4 The diagram shown is a flowchart illustrating an artificial intelligence-based power load forecasting method according to an embodiment of the present invention. In this embodiment, the artificial intelligence-based power load forecasting method includes: Historical load data, real-time meteorological data, calendar information, and power grid event data are collected from the power marketing system, meteorological center, smart meters, and distributed energy terminals, respectively. The historical load data, real-time meteorological data, calendar information, and power grid event data are partitioned and stored in the parallel chain to obtain the partitioned storage result; In the parallel chain, federated features are extracted from the partitioned storage results to obtain federated features, and the various federated features are fused to obtain spatiotemporal fusion features; The spatiotemporal fusion features are input into a preset hybrid expert model, and the predicted power load corresponding to the spatiotemporal fusion features is output through the hybrid expert model. Generate attribution auxiliary indices for the predicted power load, and use the attribution auxiliary indices to calibrate the predicted power load to the final power load.

[0061] In the several embodiments provided by this invention, it should be understood that the provided systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0062] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An artificial intelligence-based power load forecasting system, characterized in that, The system for predicting power load based on artificial intelligence includes: a data acquisition module, a data storage module, a feature fusion module, a load prediction module, and a load calibration module; The data acquisition module is used to collect historical load data, real-time meteorological data, calendar information, and power grid event data from the power marketing system, meteorological center, smart meters, and distributed energy terminals, respectively. The data storage module is used to partition and store the historical load data, the real-time meteorological data, the calendar information, and the power grid event data in the parallel chain to obtain the partition storage result; The feature fusion module is used to extract federated features from the partitioned storage results in the parallel chain to obtain federated features, and to fuse the various federated features to obtain spatiotemporal fusion features. The load forecasting module is used to input the spatiotemporal fusion features into a preset hybrid expert model, so as to output the predicted power load corresponding to the spatiotemporal fusion features through the hybrid expert model; The load calibration module is used to generate attribution auxiliary indicators for the predicted power load and to calibrate the predicted power load to the final power load using the attribution auxiliary indicators.

2. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The data collected from the electricity marketing system, meteorological center, smart meters, and distributed energy terminals includes historical load data, real-time meteorological data, calendar information, and power grid event data, including: Collect hourly electricity consumption by industry segment within historical time periods from the aforementioned electricity marketing system; The hourly electricity consumption of the aforementioned industries is used as historical load data; Temperature data, precipitation data, atmospheric pressure data, and satellite meteorological data are collected from the meteorological center. The temperature data, precipitation data, atmospheric pressure data, and satellite meteorological data are used as real-time meteorological data. Obtain calendar information from the power company's calendar data service; The first power grid event is obtained from the power company's operation and maintenance system; Collect second grid events transmitted back from the smart meters and the distributed energy terminals; The first power grid event and the second power grid event are used as power grid event data.

3. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The step of partitioning and storing the historical load data, real-time meteorological data, calendar information, and power grid event data in the parallel chain to obtain the partitioned storage result includes: Target parallel chains are defined for the historical load data, the real-time meteorological data, the calendar information, and the power grid event data; In each node of the target parallel chain, the corresponding target type data in the historical load data, the real-time meteorological data, the calendar information, and the power grid event data are preprocessed to obtain preprocessed data; In each target parachain, a pre-defined lightweight consensus mechanism is used to verify each pre-processed data and obtain the data verification results. When the data verification result is successful, the preprocessed data fragments are stored in the nodes of the target parallel chain using a partitioned storage method to obtain the original data fragments. The preprocessed data, which is sharded and stored in the nodes of each target parachain, is processed by a hash function to generate a hash digest corresponding to the preprocessed data. The hash digest is synchronized to the main chain through a cross-chain protocol between the parachain and the main chain to construct a hash index matrix corresponding to the hash digest; The original data fragments and the hash index matrix are used as the partition storage results; The target parallel chains include weather chains, load chains, and event chains.

4. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The step of extracting federated features from the partitioned storage results in the parallel chain to obtain federated features includes: Obtain the meteorological chain, load chain, and event chain from the parallel chain; Based on the partitioned storage results, temperature sensitivity coefficients and extreme weather event tags are extracted from the meteorological chain; Based on the partitioned storage results, wavelet transform is used to decompose the historical load data in the load chain to obtain high-frequency and low-frequency components. The instantaneous amplitude and frequency of the historical load data were calculated using the Hilbert-Huang transform. Based on the partitioned storage results, the data in the event chain is encoded as a time window influence factor; The temperature sensitivity coefficient, the extreme weather event label, the high-frequency component, the low-frequency component, the instantaneous amplitude, the frequency, and the time window influence factor are used as federal features.

5. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The fusion of various federated features yields spatiotemporal fusion features, including: An attention mechanism is used in the main chain to allocate attention weights for each federated feature; Based on the attention weights, smart contract technology is used in the main chain to aggregate various federated features into spatiotemporal fusion features.

6. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The step of outputting the predicted power load corresponding to the spatiotemporal fusion features through the hybrid expert model includes: Obtain the load dimension features, meteorological dimension features, and event dimension features from the spatiotemporal fusion features; The load dimension features, the meteorological dimension features, and the event dimension features are respectively input into the large model navigation layer, the LSTM expert model, and the Prophet expert model in the hybrid expert model; The long-term load data, meteorological corrections, and event offsets output by the large model navigation layer, the LSTM expert model, and the Prophet expert model are obtained respectively. Based on the meteorological correction and the event offset, the long-term load data is corrected to the predicted power load using the following formula: in, Indicates the predicted electricity load. This represents long-term load data. This indicates the amount of weather correction. Indicates the event offset. Represents the meteorological correlation coefficient. This represents the correlation coefficient of the event.

7. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The generation of the attribution auxiliary index for the predicted power load includes: Calculate the temperature contribution corresponding to the predicted power load; Calculate the event impact value corresponding to the predicted power load; The temperature contribution and the event impact value are used as attribution auxiliary indicators.

8. The artificial intelligence-based power load forecasting system as described in claim 1, characterized in that, The step of calibrating the predicted power load to the final power load using the attribution auxiliary index includes: When the temperature contribution in the attribution auxiliary indicators is greater than a preset temperature threshold and the event impact value in the attribution auxiliary indicators is greater than a preset event threshold, the original parameter data corresponding to the predicted power load is corrected to obtain corrected collected data. The predicted power load is calibrated to the final power load using the corrected collected data.

9. The artificial intelligence-based power load forecasting system as described in claim 8, characterized in that, The step of correcting the original parameter data corresponding to the predicted power load to obtain corrected collected data includes: Initiate a data completion request to the parallel chain where the data supplier corresponding to the original parameter data is located; The data completion request retrieves the completed data from the parachain where the data provider is located. The supplementary data is added to the original parameter data to obtain the corrected acquisition data.

10. A power load forecasting method based on artificial intelligence, characterized in that, The method includes: Historical load data, real-time meteorological data, calendar information, and power grid event data are collected from the power marketing system, meteorological center, smart meters, and distributed energy terminals, respectively. The historical load data, real-time meteorological data, calendar information, and power grid event data are partitioned and stored in the parallel chain to obtain the partitioned storage result; In the parallel chain, federated features are extracted from the partitioned storage results to obtain federated features, and the various federated features are fused to obtain spatiotemporal fusion features; The spatiotemporal fusion features are input into a preset hybrid expert model, and the predicted power load corresponding to the spatiotemporal fusion features is output through the hybrid expert model. Generate attribution auxiliary indices for the predicted power load, and use the attribution auxiliary indices to calibrate the predicted power load to the final power load.

Images