A method for constructing an accurate power load prediction model using big data

Abstract

The present invention discloses a method for constructing an accurate prediction model of electric load using big data, which relates to the technical field of accurate prediction of electric load, and comprises the following steps: first, through smart meters, sensors and other Internet of Things devices, multi-source data such as power operation, user information, and weather are collected every 15 minutes, abnormal data are cleaned with rule algorithms, encrypted and transmitted to a distributed storage center; then, information is retrieved from the stored data, and features strongly correlated with the load are screened and normalized by the Pearson correlation coefficient; then, a deep learning model combining LSTM and CNN is constructed, hyperparameters are set, and data sets are divided proportionally for training, so that the RMSE index of the test set is controlled within 0.1, and an accurate prediction model is obtained for subsequent load prediction. The present invention can accurately predict electric load, help the power system to operate stably, make power generation plans more reasonable, reduce energy waste, improve power supply reliability, and also reduce costs for enterprises and provide users with a better experience.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a method for constructing an accurate power load forecasting model using big data. Background Art

[0002] In today's society, the stable operation of power systems plays a crucial role in economic development and the normal functioning of social life. With the acceleration of industrialization and urbanization, electricity demand has experienced explosive growth, and fluctuations in electricity load have become increasingly complex. Traditional power load forecasting methods are no longer able to meet the requirements of refined management of modern power systems.

[0003] On the one hand, early power load forecasting relied heavily on simple statistical analysis of historical data, such as linear regression methods, which only considered roughly linear trends in electricity consumption over time. However, real-world power loads are influenced by numerous factors that defy simple linear relationships. Seasonal fluctuations in temperature and humidity can significantly alter the electricity consumption patterns of residents and businesses. Cooling equipment runs full throttle during summer heat, while heating appliances are frequently used during winter cold. These nonlinear factors lead to significant errors in traditional linear forecasting models.

[0004] Furthermore, previous forecasting methods failed to account for the diversity of user behavior. Users across different industries experience peak and off-peak seasons for industrial production, with electricity loads fluctuating significantly depending on orders and production process adjustments. Commercial users experience spikes in electricity demand during holidays and promotional events. Residential users have different lifestyles, such as those who work from 9 to 5 and those who work from home, with distinct peak electricity usage periods. Traditional methods are unable to accurately capture these complex user-side electricity usage characteristics, resulting in forecasts that deviate from reality.

[0005] Furthermore, in terms of data collection, previous technologies were fragmented and inefficient. Smart meters were not yet widespread, relying heavily on manual reading. Data updates were infrequent, making it difficult to obtain real-time electricity usage data. Furthermore, data was isolated between different departments and systems, and meteorological data, user information, and power operation data could not be effectively integrated, making it impossible to provide a comprehensive and accurate data source for load forecasting.

[0006] From a technical perspective, traditional forecasting models have limited computing power and struggle to process massive amounts of multi-source data. Emerging technologies like deep learning have been slow to penetrate the power sector. Faced with the deluge of data in the big data era, the lack of efficient data storage, processing, and model building mechanisms hinders the ability to deeply mine the value of data. This significantly reduces the accuracy and timeliness of power load forecasts, severely hindering the optimization of power system scheduling, energy conservation and loss reduction, and improved power supply reliability. New, accurate forecasting solutions based on big data are urgently needed to address this issue. Summary of the Invention

[0007] The present invention proposes a method for constructing an accurate power load prediction model using big data to solve the problems mentioned in the above-mentioned prior art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing an accurate power load prediction model using big data, comprising the following steps:

[0009] Data collection steps: Use IoT devices to collect real-time operation data at each node of the power system at intervals of 15 minutes, use cleaning algorithms to remove erroneous data, and then transmit the processed data to the data storage link through a secure encrypted channel; suppose the data transmission packet loss rate formula is , the packet loss rate is required to be controlled below 0.1%, where LP represents the number of lost packets, and TP represents the total number of packets;

[0010] Data storage steps: Use a distributed storage architecture to classify and store the collected data. The Ceph storage system uses the CRUSH algorithm for data distribution. The retrieval efficiency improvement formula is: , it is expected that the value is not less than 80%, where ORT represents the original retrieval time and NRT represents the new retrieval time.

[0011] Feature engineering steps: retrieve information from the stored data, and filter out features that are strongly correlated with power load by calculating the Pearson correlation coefficient. The feature correlation strength formula is: , select features with an absolute value greater than 0.5 so that the data value range is between [0,1]. After processing, the data is transmitted to the model construction stage;

[0012] Model construction steps: Build a prediction model based on deep learning, a hybrid model that combines the long short-term memory network LSTM with the convolutional neural network CNN, where LSTM is used to process time series data and CNN is used to extract hidden patterns in the feature space. During the training process, the model convergence speed formula is set to , it is expected to reach convergence within 80% of the training rounds, where TE represents the total training rounds of TotalEpochs and CE represents the current training round of CurrentEpoch; the model is trained using the training set until the root mean square error RMSE of the model on the test set is controlled within 0.1. The RMSE calculation formula is ,in is the true value, is the predicted value, n is the number of samples;

[0013] Model evaluation and update steps: Use the collected data to evaluate the model, and use the mean absolute error (MAE) and mean absolute percentage error (MAPE) indicators to measure the performance of the model. The mean absolute percentage error (MAPE) indicator is also used to measure the performance of the model. The MAE calculation formula is: , the MAPE calculation formula is ,in is the true value, is the predicted value, n is the number of samples;

[0014] Forecast application steps: Deploy the constructed accurate forecast model to the power dispatching center to predict the power load at different time periods in the future; present the forecast results to the dispatchers in the form of visual charts, and the dispatchers will formulate power generation plans and power allocation plans based on the forecast results.

[0015] Furthermore, it also includes:

[0016] Data fusion verification step: After the data collection step is completed, the data obtained from different data sources are fused. For data with the same attributes, the weighted average method is used for integration. This includes assigning weights to the current data of the same area collected from different smart meters according to the meter accuracy. The weight calculation formula is: ,in The accuracy of the meter is determined by using a data consistency check algorithm.

[0017] Furthermore, it also includes:

[0018] Feature dimension reduction optimization step: After the feature engineering step has screened the features, the principal component analysis (PCA) algorithm is used to reduce the feature dimension. The variance contribution rate threshold of the retained principal component is set to 90%. The formula for the principal component retention rate is: ,in is the variance of the th principal component, k is the number of retained principal components, and n is the number of original features. The random forest algorithm is then used to screen the reduced-dimensional features again to remove redundant features.

[0019] Furthermore, it also includes:

[0020] Model adaptive adjustment step: During the model building step, if the model shows signs of overfitting after a certain number of training rounds, including a continuous increase in the validation set loss value, reduce the learning rate and increase the L2 regularization coefficient. The adjustment coefficient formula is: The learning rate adjustment range is expected to be between 50%-80%, and the L2 regularization coefficient increase range is expected to be between 20%-50%, where OV represents the original value of OriginalValue and NV represents the new value of NewValue; on the contrary, if underfitting occurs, the learning rate is increased so that the model can adapt to different data distributions.

[0021] Furthermore, in the model evaluation and update steps, a user feedback collection mechanism is constructed, feedback information is quantified into weights, and integrated into the model performance evaluation index system. The feedback weight calculation formula is set as , where SS represents the SeverityScore severity score, IS represents the ImportanceScore importance score, and TS represents the TotalScore total score. The proportion of model evaluation indicators is adjusted by this weight.

[0022] Furthermore, in the prediction application step, when presenting the prediction results to the dispatcher, interactive operations are supported and a comparison function is provided.

[0023] Further, including:

[0024] Data acquisition and transmission subsystem: responsible for executing data acquisition steps and data storage steps;

[0025] Model training management subsystem: assumes the responsibilities of the feature engineering step, model building step, and model evaluation and update step in claim 1, and builds, optimizes, and maintains the prediction model;

[0026] Prediction result application subsystem: corresponds to the prediction application step in claim 1, and provides the prediction results to the power dispatcher in a visual form for use in formulating power dispatch strategies.

[0027] Furthermore, the data storage between the data storage center in the data acquisition and transmission subsystem and the model training management subsystem adopts a high-speed fiber channel with a bandwidth of not less than 10Gbps. The transmission delay reduction formula is: , it is expected that the value is not less than 60%, where OD represents the original delay and ND represents the new delay.

[0028] Furthermore, when the model training management subsystem encounters insufficient computing resources during model construction, it automatically connects to the cloud computing platform and optimizes resource allocation. The cost reduction rate formula is: , requiring the value to reach or exceed 20%, where OC represents the Original Cost and NC represents the New Cost.

[0029] Furthermore, the prediction result application subsystem uses virtual reality (VR) technology when presenting the prediction results, allowing power dispatchers to intuitively experience the trend of power load changes. Suppose the decision efficiency improvement formula is , it is expected that the value is not less than 30%, where ODT represents the original decision time and NDT represents the new decision time.

[0030] Compared with the existing technology, the beneficial effects of the present invention are:

[0031] First, for power system operations, high-precision load forecasting facilitates precise power generation planning, avoiding energy waste caused by overgeneration or power outages caused by undergeneration, optimizing power resource allocation, and reducing operating costs. Accurate forecasts allow dispatchers to pre-emptively schedule unit starts and stops, adjust power output, and ensure a balanced supply and demand, improving system stability.

[0032] Secondly, the economic benefits for power companies are significantly improved. By accurately understanding load fluctuations and rationally scheduling equipment maintenance periods, the impact of power outages on users is reduced, service quality is improved, and customer satisfaction is enhanced, thereby consolidating market share. At the same time, unnecessary standby generation capacity is reduced, investment costs are lowered, and capital utilization is improved.

[0033] Furthermore, from the perspective of energy conservation and emission reduction, accurate prediction can help avoid overproduction of energy and transmission losses, helping to achieve carbon peak and carbon neutrality goals. Based on the prediction results, the power grid flow distribution can be optimized to reduce line losses and improve energy efficiency.

[0034] Finally, it provides users with a high-quality electricity experience. Stable power supply reduces the risk of damage to electrical equipment due to voltage fluctuations, ensuring production continuity for businesses and protecting residents from power outages, creating a favorable electricity environment for both production and daily life. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic block diagram of a method for constructing an accurate power load prediction model using big data proposed in the present invention;

[0036] Figure 2 This is a schematic block diagram of a system for building an accurate power load prediction model using big data proposed by the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0039] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] Reference Figure 1-2 A method for constructing an accurate power load prediction model using big data, characterized by comprising the following steps:

[0041] Data collection steps: Utilize smart meters, sensors and other IoT devices to collect real-time operating data such as voltage, current, and power factor at high frequency at intervals of 15 minutes at each node in the power system. At the same time, establish connections with the customer information system of the power company and the meteorological department database to obtain multi-source data such as user type, electricity usage habits, weather conditions, temperature, humidity, and wind speed. Smart meters use high-precision chips with a voltage measurement accuracy of ±0.5%, a current accuracy of ±0.2%, and a power factor accuracy of ±0.01. Use a rule-based cleaning algorithm to remove duplicate, erroneous, and obviously abnormal data. For example, remove data with voltage values ​​that exceed the normal range of ±15%. Then, transmit the processed data to the data storage link through a secure encrypted channel. Assume that the data transmission packet loss rate formula is , requiring the packet loss rate to be controlled below 0.1% to ensure complete data transmission, where LP stands for LostPackets (number of lost packets) and TP stands for TotalPackets (total number of packets).

[0042] Data storage: To efficiently store and manage this massive amount of data, we adopt a distributed storage architecture. The Ceph distributed file system is an excellent choice. Ceph's unique design allows for indexing based on multiple dimensions, including data type, location, and time. For example, data type can be categorized into structured data (such as table data in databases), semi-structured data (such as XML and JSON formats), and unstructured data (such as images, videos, and documents). Geographical data can be divided based on the geographic location of data generation, which is particularly important for data management in multinational enterprises or those with multiple branches. Time data can be indexed based on the chronological order of data generation, facilitating the management and query of data across different time periods. This multi-dimensional indexing allows for categorized storage of collected data, ensuring fast retrieval within 2 seconds when processing data query requests. This fast retrieval capability is crucial for business scenarios requiring instant data access, such as real-time data analysis in financial transactions and user behavior analysis on e-commerce platforms. The Ceph storage system uses the CRUSH algorithm for data distribution. During actual data storage, the system intelligently distributes data by comprehensively considering factors such as storage device performance and network topology. Specifically, for storage device performance, the system monitors the read and write speed, storage capacity, I / O load and other indicators of each storage node, and prioritizes data allocation to nodes with better performance and lower load. For network topology, the system considers factors such as network latency and bandwidth between nodes to ensure data transmission efficiency during storage and retrieval. In order to measure the improvement in retrieval efficiency, the retrieval efficiency improvement formula is: , and we expect this value to be no less than 80%, where ORT represents Original Retrieval Time and NRT represents New Retrieval Time. To ensure data integrity and reliability, we also implement a data backup strategy. In practice, we use a combination of regular full backups and real-time incremental backups. Regular full backups involve performing a complete backup of all data at fixed intervals (such as weekly or monthly), ensuring that the complete data state at a certain point in time can be restored in the event of a major data disaster. Real-time incremental backups record and back up data changes in real time (such as additions, modifications, and deletions). This reduces the amount of backup data while ensuring the latest state of the data is protected. This dual-insurance backup strategy maximizes the security and availability of data during storage.

[0043] Feature engineering steps: retrieve the required information from the stored data, use statistical methods and professional knowledge in the power field to calculate the Pearson correlation coefficient, and filter out features that are strongly correlated with power load, such as temperature during a specific period of time, historical peak power consumption of different user types, etc. Suppose the feature correlation strength formula is , select features with absolute values ​​greater than 0.5, normalize the selected features so that the data value range is between [0,1], and transmit the data to the model construction link after processing.

[0044] Model construction steps: Build a prediction model based on deep learning, specifically a hybrid model that combines a long short-term memory network (LSTM) with a convolutional neural network (CNN). LSTM is good at processing time series data, and CNN is used to extract hidden patterns in the feature space. Set the model's hyperparameters, such as setting the number of LSTM units to 128 and the learning rate to 0.001. Use the adaptive moment estimation (Adam) optimization algorithm and divide the historical load data into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10%. During the training process, set the model convergence rate formula to be , it is expected to reach convergence within 80% of the training rounds, where TE represents Total Epochs (total training rounds) and CE represents Current Epoch (current training round); the model is trained on the training set and adjusted and optimized on the validation set until the root mean square error (RMSE) of the model on the test set is controlled within 0.1, thus obtaining an accurate prediction model. The RMSE calculation formula is ,in is the true value, is the predicted value, and n is the number of samples.

[0045] Model evaluation and update steps: Regularly, such as weekly, use the latest collected data to evaluate the model. In addition to the root mean square error (RMSE), the mean absolute error (MAE), mean absolute percentage error (MAPE) and other indicators are also used to measure the performance of the model. The MAE calculation formula is , the MAPE calculation formula is When the model's performance degrades, for example, when the RMSE increases by more than 10%, the model's update mechanism is triggered, recent data is recollected, and the feature engineering and model building steps are repeated to optimize the model and ensure that the accuracy of the predictions continues to improve.

[0046] Predictive Application Steps: Predictive applications play a crucial role in the operation and management of power systems. First, a carefully constructed, accurate predictive model must be deployed to the power dispatch center. Building this predictive model is a complex and rigorous process, involving extensive data collection and analysis. This data source includes historical operating data from various aspects of the power system, such as power generation at power plants, load on transmission lines, and voltage and current data at substations. The impact of external factors on power demand, such as weather conditions, seasonal variations, and holidays, is also considered. Advanced data mining techniques and machine learning algorithms are used to conduct in-depth analysis and training of this data, resulting in a highly accurate predictive model. Once successfully deployed at the power dispatch center, this predictive model possesses powerful capabilities. Its core capability is the ability to receive real-time data on the current operation of the power system. This means it establishes close connections with various monitoring devices and data acquisition systems within the power system, acquiring the latest data every minute and every second, including but not limited to real-time power generation at each power plant, real-time load conditions on each transmission line, and real-time voltage and current data at each substation. Based on this real-time data, the predictive model can then predict power load at different time periods in the future. For example, it can accurately predict hourly power load changes over the next 24 hours and power demand trends for each time period over the next 48 hours. These forecasts are invaluable decision-making tools for power dispatchers. To make these forecasts more intuitive, easier to understand, and use, dispatchers are presented with visual charts. These charts come in a variety of types, including line charts and bar charts. Line charts clearly illustrate power load trends over time. Dispatchers can determine whether power demand is increasing or decreasing, and the magnitude of the change, by observing the direction of the line. Bar charts can be used to compare power load over different time periods, such as the same time period on different dates or different times of the day, helping dispatchers quickly identify peak and trough periods. After receiving these intuitive forecast results, dispatchers use them to formulate appropriate power generation plans and power allocation strategies. When formulating power generation plans, they take into account factors such as the power generation capacity of each power station, fuel reserves, and equipment maintenance schedules. For example, if a significant increase in electricity demand is predicted for a certain period in the future, dispatchers will notify thermal power plants in advance to increase coal reserves to ensure sufficient power output during the high-demand period. For hydroelectric power plants, turbine operation schedules will be appropriately scheduled based on reservoir water levels and water inflow forecasts. Regarding power allocation, dispatchers will adjust the direction and volume of power transmission along transmission lines based on the distribution of power loads.For example, if a sharp increase in electricity demand is predicted for a specific area, the power distribution of nearby transmission lines will be adjusted promptly, sending more electricity from areas with more abundant power supply to that area to avoid power shortages. At the same time, power surpluses must also be prevented, as they can lead to unnecessary operation of power generation equipment, increasing energy waste and operating costs. This comprehensive set of predictive application steps can effectively ensure the stable operation of the power system, ensuring that user power needs are met while achieving the rational allocation and efficient utilization of power resources.

[0047] The present invention also includes a data fusion verification step: after the data collection step is completed, the data obtained from different data sources are fused, and the data with the same attributes are integrated using a weighted average method. For example, for the current data of the same area collected from different smart meters, weights are assigned according to the meter accuracy. The weight calculation formula is: ,in To ensure the accuracy of each electricity meter, a data consistency verification algorithm is used. At the same time, a data consistency verification algorithm is used to check whether there are logical contradictions in the fused data. If it is found that the temperature in a certain period is inconsistent with the usual climate laws of that period, the data source is traced back to re-collect data in time to ensure data accuracy and provide a reliable foundation for subsequent model training.

[0048] The present invention also includes a feature dimensionality reduction optimization step: after the feature engineering step has screened the features, the principal component analysis (PCA) algorithm is used to reduce the dimensionality of the features, converting the original high-dimensional and complex feature space into a low-dimensional and concise feature representation, and setting the variance contribution rate threshold of the principal component to 90%. While reducing the computational complexity, the information valuable to the power load forecast is retained to the greatest extent. Assume that the principal component retention rate formula is ,in is the variance of the th principal component, k is the number of retained principal components, and n is the number of original features. The random forest algorithm is then used to screen the reduced-dimensional features again to remove redundant features and further improve the quality of the model input.

[0049] The present invention also includes a model adaptive adjustment step: during the model building step, after a certain number of training rounds (such as 10 epochs), the hyperparameters are dynamically adjusted according to the performance of the current model on the validation set. If the model shows signs of overfitting, such as a continuous increase in the validation set loss value, the learning rate is reduced and the L2 regularization coefficient is increased. The adjustment coefficient formula is set to be The learning rate adjustment range is expected to be between 50%-80%, and the L2 regularization coefficient increase range is expected to be between 20%-50%, where OV represents OriginalValue and NV represents NewValue. On the contrary, if underfitting occurs, the learning rate should be appropriately increased so that the model can adapt to different data distributions and continuously improve the prediction performance.

[0050] In the present invention, in the model evaluation and update step, a user feedback collection mechanism is constructed to collect feedback on the prediction results from power dispatchers and power users, such as whether the prediction accuracy meets the actual needs, whether there is a large deviation in the load prediction results for a certain period of time that is not captured by the model, and the feedback information is quantified into weights and integrated into the model performance evaluation index system to make the model update more suitable for actual application scenarios. Suppose the feedback weight calculation formula is , where SS represents Severity Score, IS represents Importance Score, and TS represents Total Score. By adjusting the proportion of model evaluation indicators through this weight, model optimization can be more targeted.

[0051] In the present invention, during the forecast application step, when presenting forecast results to dispatchers, a rich set of interactive functions is provided. Dispatchers can view detailed load forecast data for different regions and user types by clicking and dragging the mouse. Specifically, when viewing data for different regions, dispatchers can use the mouse to click on a specific geographical area on the visual interface, such as various districts in a city, different industrial parks, commercial areas, or residential areas. The system will immediately present detailed power load forecast data for that area. This data includes peak and valley power demand values, as well as power load trends, for different time periods in the future. At the same time, by dragging the mouse, dispatchers can flexibly switch between different regions, facilitating a comprehensive and detailed analysis of the entire power supply area. When viewing load forecast data for different user types, the system can distinguish between different types, such as industrial users, commercial users, and residential users. When the dispatcher clicks the corresponding user type option, the system will display the power usage forecast for that type of user for different time periods in the future. For example, for industrial users, detailed power demand forecasts for businesses of varying sizes, from large factories to small workshops, are displayed, including power consumption during peak and off-peak periods. For commercial users, power load forecasts for shopping malls, office buildings, and hotels are presented, taking into account the impact of business hours, holidays, and other factors on power demand. For residential users, power usage forecasts for different time periods are calculated based on residents' daily habits and seasonal variations. The system also offers a powerful comparison feature. This feature allows users to take snapshots of current forecast results and compare them with the same period in history and the same period last week. When comparing with historical data, the system automatically retrieves power load data for the same time period over the past few years and presents it in intuitive charts. For example, a line chart displays the current forecasted power load curve against the power load curves for the same date in previous years, allowing dispatchers to clearly identify power demand trends and identify any anomalies. When comparing with the same period last week, the system quickly retrieves power load data for the same period last week and compares it with the current forecast. This comparison helps dispatchers understand short-term fluctuations in power demand, particularly when factoring in cyclical factors (such as differences in power demand between weekdays and weekends). This allows them to better formulate appropriate power generation plans and dispatching strategies. These rich interactive operations and comparison features help dispatchers make more accurate decisions, ensuring stable power system operation, rationally allocating power resources, and avoiding power shortages or oversupply.

[0052] In the present invention, the data acquisition and transmission subsystem is responsible for executing the data acquisition steps and data storage steps to ensure data acquisition, transmission and storage;

[0053] Model training management subsystem: It is responsible for feature engineering, model building, and model evaluation and update steps, and builds, optimizes, and maintains the prediction model.

[0054] Prediction result application subsystem: Corresponding to the prediction application steps, the prediction results are provided to the power dispatchers in a visual form for formulating power dispatch strategies.

[0055] In the present invention, the data storage center in the data acquisition and transmission subsystem and the data storage center in the data acquisition and transmission subsystem and the data transmission between the model training management subsystem adopt a high-speed fiber channel with a bandwidth of not less than 10Gbps, which reduces data snapshots, reduces data transmission delay, and improves model training efficiency. Assume that the transmission delay reduction formula is , it is expected that the value is not less than 60%, where OD represents OriginalDelay and ND represents NewDelay.

[0056] In the present invention, when the model training management subsystem encounters insufficient computing resources during model construction, it automatically accesses cloud computing platforms, such as Alibaba Cloud and Tencent Cloud, and utilizes elastic computing resources to ensure smooth model construction and optimization. At the same time, it optimizes resource allocation and reduces cloud computing costs by more than 20%. Assume that the cost reduction rate formula is , requiring the value to reach or exceed 20%, where OC represents Original Cost and NC represents New Cost.

[0057] In this invention, the forecast result application subsystem uses virtual reality (VR) or augmented reality (AR) technology to present forecast results, providing power dispatchers with an immersive load forecast display. In practice, when power dispatchers need to view load forecast results, they are no longer limited to traditional two-dimensional charts and data tables. Instead, they wear VR equipment to enter a virtual power system operation scenario. In this virtual scenario, the entire power system architecture is presented in three dimensions, including components such as power stations, substations, and transmission lines. Dispatchers can observe every detail of the power system in a 360-degree perspective, as if they were actually in the actual power system operation site. Each power station is accurately modeled in the virtual scenario, its appearance and scale consistent with its actual power station. Different colors and dynamic effects are used to indicate the current operating status of the power station. For example, a power station with sufficient power generation may appear bright blue, while a power station nearing full capacity may appear red. The substation's voltage conversion and power distribution functions are displayed in the virtual scenario. Dispatchers can view detailed data such as the current voltage level, power input, and output by approaching the substation. Transmission lines crisscross the virtual scene, just like a real power grid. The brightness and color of the lines change to reflect the flow of electricity. For example, lines with heavy loads may appear thicker and brighter, while lines with light loads may appear thinner and darker.

[0058] To display load forecast results, the virtual scene uses brightness and dynamic effects to indicate future load trends. For example, if a region's power demand is predicted to increase significantly, the region's brightness will gradually brighten in the virtual scene, and dynamic arrows or ripples may appear to indicate the upward trend in power load. Dispatchers can view power load forecasts for different regions by walking or flying through the virtual scene. This immersive experience allows them to more intuitively understand the temporal and spatial distribution and changing patterns of power load. If augmented reality (AR) technology is used, dispatchers can wear AR glasses and see a virtual model of the power system and forecast data superimposed on the real scene in a real environment. For example, AR technology can project a virtual power system operating scenario and load forecast results directly onto the large screen in the power dispatch center. Dispatchers can view the virtual forecast data and model while viewing the actual geographic map or power system layout diagram. This combination of virtual and real allows dispatchers to enjoy the convenience of advanced technology in a familiar work environment, more intuitively understand power load trends, and make more accurate and timely power dispatch decisions. Let the decision efficiency improvement formula be: , it is expected that the value is not less than 30%, where ODT represents OriginalDecisionTime and NDT represents NewDecisionTime.

[0059] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for constructing an accurate power load prediction model using big data, characterized in that: The following steps are involved: Data collection steps: Use IoT devices to collect real-time operating data at various nodes in the power system, apply cleaning algorithms to remove erroneous data, and then transmit the processed data to the data storage link through a secure encrypted channel; Data storage steps: A distributed storage architecture is used to classify and store the collected data. The Ceph storage system uses the CRUSH algorithm for data distribution. Feature engineering steps: retrieve information from the stored data, and filter out features that are strongly correlated with power load by calculating the Pearson correlation coefficient. The feature correlation strength formula is: , select features with an absolute value greater than 0.5 so that the data value range is between [0,1]. After processing, the data is transmitted to the model construction stage; Model construction steps: Build a prediction model based on deep learning, a hybrid model that combines the long short-term memory network LSTM with the convolutional neural network CNN. During the training process, the model convergence speed formula is set as , where TE represents the total training rounds of TotalEpochs, and CE represents the current training round of CurrentEpoch. The model is trained using the training set, and the RMSE calculation formula is: ,in is the true value, is the predicted value, n is the number of samples; Model evaluation and update step: Use the collected data to evaluate the model, and use the mean absolute error (MAE) and mean absolute percentage error (MAPE) indicators to measure the performance of the model. The mean absolute percentage error (MAPE) indicator is also used to measure the performance of the model; Prediction application steps: Deploy the constructed accurate prediction model to the power dispatching center; present the prediction results to the dispatchers in the form of visual charts.

2. The method for constructing an accurate power load prediction model using big data according to claim 1, characterized in that: Also includes: Data fusion verification step: After the data collection step is completed, the data obtained from different data sources are fused. For data with the same attributes, the weighted average method is used for integration. This includes assigning weights to the current data of the same area collected from different smart meters according to the meter accuracy. The weight calculation formula is: ,in The accuracy of the meter is determined by using a data consistency check algorithm.

3. The method for constructing an accurate power load prediction model using big data according to claim 1, characterized in that: Also includes: Feature dimension reduction optimization step: After the feature engineering step has screened the features, the principal component analysis (PCA) algorithm is used to reduce the feature dimension. The variance contribution rate threshold of the retained principal component is set to 90%. The formula for the principal component retention rate is: ,in is the variance of the th principal component, k is the number of retained principal components, and n is the number of original features. The random forest algorithm is then used to screen the reduced-dimensional features again to remove redundant features.

4. The method for constructing an accurate power load prediction model using big data according to claim 1, characterized in that: Also includes: Model adaptive adjustment step: During the model building step, if the model shows signs of overfitting after a certain number of training rounds, including a continuous increase in the validation set loss value, reduce the learning rate and increase the L2 regularization coefficient. The adjustment coefficient formula is: The learning rate adjustment range is expected to be between 50%-80%, and the L2 regularization coefficient increase range is expected to be between 20%-50%, where OV represents the original value of OriginalValue and NV represents the new value of NewValue; on the contrary, if underfitting occurs, the learning rate is increased so that the model can adapt to different data distributions.

5. The method for constructing an accurate power load prediction model using big data according to claim 1, characterized in that: In the model evaluation and update steps, a user feedback collection mechanism is constructed, feedback information is quantified into weights, and integrated into the model performance evaluation index system. The feedback weight calculation formula is set as , where SS represents the SeverityScore severity score, IS represents the ImportanceScore importance score, and TS represents the TotalScore total score. The proportion of model evaluation indicators is adjusted by this weight.

6. The method for constructing an accurate power load prediction model using big data according to claim 1, characterized in that: During the forecast application step, when presenting forecast results to dispatchers, interactive operations are supported and comparison functions are provided.

7. A system for implementing the method for constructing a power load accurate prediction model using big data as described in any one of claims 1 to 6, characterized in that: include: Data acquisition and transmission subsystem: responsible for executing the data acquisition step and data storage step described in claim 1; Model training management subsystem: undertakes the responsibilities of the feature engineering step, model building step, and model evaluation and update step described in claim 1, and builds, optimizes, and maintains the prediction model; Prediction result application subsystem: corresponds to the prediction application step described in claim 1, and provides the prediction results to the power dispatcher in a visual form for use in formulating power dispatch strategies.

8. The system for constructing a power load accurate prediction model using big data according to claim 7, characterized in that: The data transmission between the data storage center in the data acquisition and transmission subsystem and the model training management subsystem adopts a high-speed fiber channel with a bandwidth of no less than 10Gbps. The transmission delay reduction formula is: , the expected calculated value is no less than 60%, where OD represents the original delay and ND represents the new delay.

9. The system for constructing a power load accurate prediction model using big data according to claim 7, characterized in that: When the model training management subsystem encounters insufficient computing resources during model construction, it automatically connects to the cloud computing platform and optimizes resource allocation. The cost reduction rate formula is: , the calculated value is required to reach or exceed 20%, where OC represents the original cost and NC represents the new cost.

10. The system for constructing a power load accurate prediction model using big data according to claim 7, characterized in that: When presenting the prediction results, the prediction result application subsystem uses virtual reality (VR) technology to enable power dispatchers to intuitively feel the trend of power load changes. The decision-making efficiency improvement formula is: , the expected calculated value is not less than 30%, where ODT represents the original decision time of OriginalDecisionTime and NDT represents the new decision time of NewDecisionTime.

Images