Ultra-short-term power load forecasting methods, program products, storage media and equipment

Abstract

This invention provides an ultra-short-term power load forecasting method, program product, storage medium, and device. The forecasting method includes: acquiring historical power load data and historical external factor data stored in a historical database to generate a historical pattern feature vector representing a typical daily load pattern; acquiring real-time power load data and real-time external factor data from the start of the day to the current time to generate a real-time dynamic feature vector representing the daily power load change trend; dynamically determining the fusion weights for fusing the historical pattern feature vector and the real-time dynamic feature vector; using the fusion weights to perform weighted fusion of the historical pattern feature vector and the real-time dynamic feature vector to obtain a fused feature vector, and outputting the ultra-short-term power load forecast result from the current time to the end of the day. The advantage of this invention is that it can achieve adaptive fusion of historical experience and real-time data by dynamically determining the fusion weights, thereby improving forecast accuracy.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to an ultra-short-term power load forecasting method, program product, storage medium and device. Background Technology

[0002] Electricity load forecasting serves as a crucial basis for grid dispatching, operation planning, and market transactions, and is fundamental to achieving the safe, stable, and economical operation of the power system. Based on the forecasting time scale, load forecasting can be categorized into long-term, medium-term, short-term, and ultra-short-term forecasts. Among these, ultra-short-term load forecasting for the remaining time of the day is particularly important for supporting real-time economic grid dispatching, rapid decision-making regarding reserve capacity, and promoting the absorption of a high proportion of fluctuating renewable energy sources.

[0003] Existing ultra-short-term load forecasting methods generally adopt a data-driven modeling approach. These methods typically build a static or only periodically updated general forecasting model based on closed historical datasets, aiming to learn the complex mapping relationship between load changes and various factors. Although such models can capture the typical patterns contained in historical data, they are essentially a fixed expression of past experience.

[0004] However, in real-time forecasting scenarios, real-time data may deviate from historical patterns, and traditional models lack a mechanism to dynamically adjust their forecasting basis. This makes it difficult to adaptively balance the contributions of real-time data and historical experience in forecasting. This limitation leads to a significant decline in the model's predictive performance and adaptability when load patterns change rapidly or unexpected events occur. Summary of the Invention

[0005] One objective of this invention is to provide an ultra-short-term power load forecasting method that can dynamically determine fusion weights to achieve adaptive fusion of historical experience and real-time data, thereby improving forecasting accuracy.

[0006] A further objective of this invention is to extract time-series evolution features and dynamic response features from real-time power load data, thereby making the generation of real-time dynamic feature vectors more accurate.

[0007] Another further objective of this invention is to reduce the computational overhead of model updates, while ensuring that the predictive model can keep up with the latest changes in the system in a timely manner, and maintain efficient and sustainable online operation capabilities.

[0008] In particular, according to a first aspect of the present invention, the present invention provides an ultra-short-term power load forecasting method, comprising:

[0009] Acquire historical power load data and historical external factor data stored in the historical database, and generate historical pattern feature vectors that characterize typical daily load patterns;

[0010] Acquire real-time power load data and real-time external factor data from the start of the day to the present, and generate a real-time dynamic feature vector characterizing the trend of power load changes on the day;

[0011] Dynamically determine the fusion weights used to fuse historical pattern feature vectors and real-time dynamic feature vectors;

[0012] The historical pattern feature vector and the real-time dynamic feature vector are weighted and fused using fusion weights to obtain a fused feature vector.

[0013] The ultra-short-term power load forecast results from the current time to the end of the day are output based on the fused feature vector.

[0014] Optionally, the steps for generating historical pattern feature vectors representing typical daily load patterns include:

[0015] Historical power load data and historical external factor data are input into the first neural network model;

[0016] The first neural network model is used to analyze the periodic variation pattern of historical power load data at different time scales, and the periodic variation pattern at each time scale is corrected based on historical external factor data, thereby generating historical pattern feature vectors.

[0017] Optionally, the step of generating a real-time dynamic feature vector representing the daily electricity load change trend includes:

[0018] Real-time power load data and real-time external factor data are input into the second neural network model;

[0019] The second neural network model is used to extract the time-series evolution characteristics of real-time power load data and the dynamic response characteristics of real-time external factor data;

[0020] By capturing the continuous trend of daily load changes based on time-series evolution characteristics and quantifying the real-time impact of external factor fluctuations on load based on dynamic response characteristics, the deviation between real-time power load data and typical daily load patterns can be identified, thereby generating a real-time dynamic feature vector.

[0021] Optionally, the step of dynamically determining the fusion weights for fusing historical pattern feature vectors and real-time dynamic feature vectors includes:

[0022] Calculate the difference between the real-time dynamic feature vector and the historical pattern feature vector;

[0023] The first weight of the historical pattern feature vector and the second weight of the real-time dynamic feature vector are generated based on the difference. The greater the difference, the higher the proportion of the second weight relative to the first weight.

[0024] Optionally, the step of generating the first weight of the historical pattern feature vector and the second weight of the real-time dynamic feature vector based on the difference includes:

[0025] The first and second weights are calculated based on the degree of difference using an attention mechanism; or

[0026] After concatenating the historical pattern feature vector with the real-time dynamic feature vector, the first weight and the second weight are obtained by mapping them through a preset neural network based on the degree of difference.

[0027] Optionally, after the step of outputting the ultra-short-term power load forecast results from the current time to the end of the day based on the fused feature vector, the method further includes:

[0028] The ultra-short-term power load forecast results, the forecast time corresponding to the ultra-short-term power load forecast results, and real-time external factor data are associated and stored in the historical database.

[0029] Optionally, after the step of outputting the ultra-short-term power load forecast results from the current time to the end of the day based on the fused feature vector, the method further includes:

[0030] In response to receiving new real-time load data, the deviation between the real-time load value of the new real-time load data and the load forecast value corresponding to the current time in the previous forecast is calculated;

[0031] When the deviation exceeds the preset threshold at multiple consecutive moments, the parameters of the model part that generates the real-time dynamic feature vector in the prediction model are adjusted using sliding window data composed of real-time power load data and corresponding real-time external factor data from recent consecutive periods.

[0032] According to a second aspect of the present invention, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the steps of the ultra-short-term power load forecasting method described above.

[0033] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the ultra-short-term power load forecasting method described above.

[0034] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the ultra-short-term power load forecasting method described above.

[0035] The ultra-short-term power load forecasting method of this invention utilizes historical power load data and historical external factor data to generate a historical pattern feature vector representing a typical daily load pattern. It then utilizes real-time power load data from the start of the day to the current time and real-time external factor data to generate a real-time dynamic feature vector representing the daily power load change trend. By dynamically determining the fusion weights between the historical pattern feature vector and the real-time dynamic feature vector, adaptive fusion of historical load patterns and real-time operating conditions can be achieved, forming a more comprehensive fused feature vector. Based on this fused feature vector, the power load forecasting result for the remaining time period of the day is output. This dynamic weight allocation mechanism based on real-time conditions effectively improves the adaptability of the forecasting model to different operating conditions and significantly enhances the accuracy of ultra-short-term load forecasting.

[0036] Furthermore, the ultra-short-term power load forecasting method of the present invention extracts the temporal evolution characteristics of real-time power load data and the dynamic response characteristics of real-time external factor data. It can capture the continuous changing trend of daily load using the temporal evolution characteristics and quantify the real-time impact of external factor fluctuations on the load using the dynamic response characteristics. Based on these two types of extracted features, the method can accurately identify the deviation between real-time power load data and typical daily load patterns, thereby generating a more accurate real-time dynamic feature vector.

[0037] Furthermore, the ultra-short-term power load forecasting method of the present invention, in response to receiving new real-time load data, calculates the deviation between the real-time load value of the new real-time load data and the load forecast value corresponding to the current moment in the previous forecast. When the deviation exceeds a preset threshold for multiple consecutive moments, the training parameters of the model generating the real-time dynamic feature vector are adjusted using a sliding window of data composed of recent consecutive periods of real-time power load data and corresponding real-time external factor data. This method only incrementally updates the model part of the forecasting model that generates the real-time dynamic feature vector, significantly reducing computational costs while enabling the forecasting model to quickly adapt to system changes, thereby maintaining the accuracy and timeliness of forecasts over the long term.

[0038] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0039] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0040] Figure 1 This is a flowchart of an ultra-short-term power load forecasting method according to an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of the step of generating historical pattern feature vectors in an ultra-short-term power load forecasting method according to an embodiment of the present invention;

[0042] Figure 3 This is a flowchart of the step of generating real-time dynamic feature vectors in an ultra-short-term power load forecasting method according to an embodiment of the present invention;

[0043] Figure 4 This is a flowchart of the model self-learning and self-correction steps in an ultra-short-term power load forecasting method according to an embodiment of the present invention;

[0044] Figure 5 This is a schematic block diagram of a prediction model according to an embodiment of the present invention;

[0045] Figure 6 This is a schematic diagram of a computer program product according to an embodiment of the present invention;

[0046] Figure 7 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention;

[0047] Figure 8 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0048] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0049] This invention provides an ultra-short-term power load forecasting method, which is mainly used to predict the load curve for the remaining time period of the day. Figure 1 This is a schematic diagram of an ultra-short-term power load forecasting method according to an embodiment of the present invention. The method generally includes:

[0050] Step S101: Obtain historical power load data and historical external factor data stored in the historical database, and generate a historical pattern feature vector representing the typical daily load pattern. The historical power load data indicates the hourly load value sequence for multiple typical days in the past (such as weekdays, weekends, and holidays); the historical external factor data indicates historical weather data for the corresponding time period, such as historical temperature, humidity, and wind speed.

[0051] Step S102: Acquire real-time power load data and real-time external factor data from the start time of the day to the current time, and generate a real-time dynamic feature vector characterizing the daily power load change trend. The start time can be 00:00 on the current day, and the current time refers to the real-time system time at which this forecast is being executed. Real-time power load data indicates the load sequence generated on the current day; real-time external factor data indicates the measured weather data and short-term weather forecast data up to the current time.

[0052] Step S103: Dynamically determine the fusion weights used to fuse historical pattern feature vectors and real-time dynamic feature vectors. In other words, the fusion weights are not fixed, but are dynamically calculated based on the difference between the real-time dynamic feature vectors and historical pattern feature vectors.

[0053] Step S104: The historical pattern feature vector and the real-time dynamic feature vector are weighted and fused using fusion weights to obtain a fused feature vector. This vector integrates information from both long-term historical patterns and short-term daily dynamics, forming a more comprehensive and accurate representation of the current load status.

[0054] Step S105: Output the ultra-short-term power load forecast results from the current time to the end time of the day based on the fused feature vector. The end time of the day can be 24:00 on the same day. The ultra-short-term power load forecast results are used to indicate the load curve forecast for the remaining time period of the day.

[0055] The method described in the above embodiments utilizes historical power load data and historical external factor data to generate a historical pattern feature vector representing a typical daily load pattern. It then utilizes real-time power load data from the start of the day to the current time and real-time external factor data to generate a real-time dynamic feature vector representing the daily power load change trend. By dynamically determining the fusion weights between the historical pattern feature vector and the real-time dynamic feature vector, adaptive fusion of historical load patterns and real-time operating conditions can be achieved, forming a more comprehensive fused feature vector. Based on this fused feature vector, the power load forecast for the remaining time period of the day is output. This dynamic weight allocation mechanism based on real-time conditions effectively improves the adaptability of the forecast model to different operating conditions and significantly enhances the accuracy of ultra-short-term load forecasting.

[0056] Figure 2This is a flowchart of the step of generating historical pattern feature vectors in an ultra-short-term power load forecasting method according to an embodiment of the present invention, as follows: Figure 2 As shown, the steps for generating historical pattern feature vectors representing typical daily load patterns may include:

[0057] Step S201: Input historical power load data and historical external factor data into the first neural network model.

[0058] Step S202: Analyze the periodic variation patterns of historical power load data at different time scales using the first neural network model. These different time scales may include daily, weekly, and seasonal time scales. Correct the periodic variation patterns at each time scale based on historical external factor data to generate historical pattern feature vectors.

[0059] Using the method described in this embodiment, robust load pattern features can be automatically and structurally extracted from massive and complex historical data. This process overcomes the limitations of traditional methods that rely on manual feature engineering or single time-series models. By automatically mining the coupling relationship between multi-scale periodicity and external factors through deep models, it provides a high-quality, highly interpretable historical reference benchmark for subsequent prediction processes. This not only enhances the model's understanding of historical patterns but also lays a reliable data foundation for the adaptive fusion of historical experience and real-time dynamics.

[0060] In one example, the generation of historical pattern feature vectors can be achieved through a historical pattern extraction module. This module is built upon a first neural network model, which can be a two-dimensional time-based network (TimesNet). By inputting historical electricity load data and historical external factor data into the historical pattern extraction module, it can output a fixed-dimensional historical pattern feature vector. This historical pattern feature vector is a dense numerical vector that internally encodes stable and repeatable typical daily load pattern information extracted from historical data, corrected for environmental factors.

[0061] In another example, the TimeNet model can be replaced by other models that can extract periodic features of time series, such as convolutional neural networks (CNNs) to extract local patterns or shape features of the load curve.

[0062] Figure 3 This is a flowchart of the step of generating real-time dynamic feature vectors in an ultra-short-term power load forecasting method according to an embodiment of the present invention, as follows: Figure 3 As shown, the steps for generating a real-time dynamic feature vector representing the daily electricity load change trend may include:

[0063] Step S301: Input the real-time power load data and real-time external factor data into the second neural network model.

[0064] Step S302: Use the second neural network model to extract the time-series evolution characteristics of real-time power load data and the dynamic response characteristics of real-time external factor data.

[0065] Step S303: Capture the continuous change trend of the daily load based on the time-series evolution characteristics, quantify the real-time impact of external factor fluctuations on the load based on the dynamic response characteristics, so as to identify the deviation between the real-time power load data and the typical daily load pattern, thereby generating a real-time dynamic feature vector.

[0066] Using the method described in this embodiment, by extracting the temporal evolution characteristics of real-time power load data and the dynamic response characteristics of real-time external factor data, the continuous changing trend of the daily load can be captured using the temporal evolution characteristics, and the real-time impact of external factor fluctuations on the load can be quantified using the dynamic response characteristics. Based on these two types of extracted features, this method can accurately identify the deviation between real-time power load data and typical daily load patterns, thereby generating a more accurate real-time dynamic feature vector.

[0067] In one example, the generation of real-time dynamic feature vectors can be achieved through a real-time dynamic capture module. This module is built upon a second neural network model, which can be a Long Short-Term Memory (LSTM) network. By inputting real-time power load data and real-time external factor data into the real-time dynamic capture module, it can output a real-time dynamic feature vector. This vector is a dynamically evolving dense numerical vector that encodes the actual temporal evolution characteristics of the load from the start of the day to the current time, the dynamic response to fluctuations in real-time external factors, and the real-time deviation information between the vector and the typical daily load pattern represented by the historical pattern feature vector, thereby capturing instantaneous changes in the load pattern caused by sudden events or unforeseen factors.

[0068] Specifically, the real-time dynamic capture module receives real-time power load data from the start of the day to the current moment and uses its internal LSTM model to process this data hour by hour, thereby extracting the dynamic evolution trajectory of load changes. For example, even if the current load value is within the normal range, the LSTM model can identify a potential situation where the load is about to exceed the typical level for the same period in history based on the continuous rapid upward trend in the previous few hours. The LSTM model simultaneously processes real-time external factor data, learning the instantaneous correlation between these external factors and load changes. For example, when the real-time solar irradiance suddenly decreases at noon due to cloud cover, the LSTM model can quantify the instantaneous impact of this change on commercial and photovoltaic power output.

[0069] By jointly analyzing these two types of features, the LSTM model can identify the deviation between the current real-time operating status and historical typical patterns. For example, on a weekday morning, if the time-series evolution features show weak load growth, while the dynamic response features show that the temperature is significantly higher than the historical average for the same period, the LSTM model will comprehensively determine that the current load level is abnormally lower than the historical expectation after considering the influence of temperature, and encode this information in the generated real-time dynamic feature vector.

[0070] In another example, the LSTM model can be replaced by other models that are good at capturing temporal dependencies. For example, Gated Recurrent Units (GRUs), Transformers (self-attention models), or Graph Attention Networks (GATs).

[0071] In some embodiments, when dynamically determining the fusion weights for fusing historical pattern feature vectors and real-time dynamic feature vectors, the weight allocation basis can be generated by calculating the difference between the real-time dynamic feature vectors and historical pattern feature vectors. The difference can be quantified using cosine distance, Euclidean distance, or other similarity measures. Then, a first weight for the historical pattern feature vector and a second weight for the real-time dynamic feature vector are generated based on the difference, wherein the greater the difference, the higher the proportion of the second weight relative to the first weight.

[0072] In other words, when the difference between real-time dynamic features and historical pattern features is large, it indicates that the current operating state deviates significantly from the historical norm. At this time, the system will assign higher weight to the real-time dynamic features of the day, making the final prediction result more dependent on the latest operating trend, thereby enhancing the model's ability to respond to sudden situations and abnormal patterns. Conversely, if the difference is small, the system will make predictions based more on stable historical patterns to maintain the robustness of the results.

[0073] In one example, an attention mechanism can be used to calculate the first and second weights based on the degree of difference. Specifically, the real-time dynamic feature vector is used as the query vector, and the historical pattern feature vector is used as both the key and value vectors. The weight allocation of the historical pattern feature vector is dynamically calculated using the scaled dot product attention formula to obtain the first and second weights. This process essentially quantifies and utilizes the degree of difference between the real-time dynamic feature vector and the historical pattern feature vector. This approach allows for more precise identification and focus on the historical pattern components most relevant to the current real-time state, thereby achieving more accurate feature fusion.

[0074] In another example, the fusion weights are calculated as follows: First, the historical pattern feature vector is concatenated with the real-time dynamic feature vector. Based on the degree of difference, a pre-defined neural network is used to map the two vectors to obtain the first and second weights. Specifically, this neural network learns to extract and utilize the degree of difference between the concatenated feature vectors, and finally outputs the corresponding first and second weights. This method implicitly learns the complex interaction relationships between features through the neural network, enabling it to adaptively generate the first and second weights.

[0075] In some embodiments, after outputting the ultra-short-term power load forecast results from the current time to the end of the day based on the fused feature vector, the ultra-short-term power load forecast results, the corresponding forecast time, and real-time external factor data can be associated and stored in a historical database. This creates a closed-loop data feedback and accumulation mechanism for the forecast results, providing a continuously updated data source for subsequent model retraining or periodic updates of historical load patterns.

[0076] Figure 4 This is a flowchart of the model self-learning and self-correction steps in an ultra-short-term power load forecasting method according to an embodiment of the present invention, as follows: Figure 4 As shown, after outputting the ultra-short-term power load forecast results, the following steps can be performed:

[0077] Step S401: In response to receiving new real-time load data, calculate the deviation between the real-time load value of the new real-time load data and the load forecast value corresponding to the current time in the previous forecast.

[0078] Step S402: When the deviation exceeds the preset threshold at multiple consecutive times, the parameters of the model part that generates the real-time dynamic feature vector in the prediction model are adjusted using sliding window data composed of real-time power load data and corresponding real-time external factor data from recent consecutive periods.

[0079] Using the above method, incremental updates are performed only on the model portion that generates real-time dynamic feature vectors (i.e., the second neural network model), while keeping the parameters of the model portion that generates historical pattern feature vectors (i.e., the first neural network model) unchanged. This significantly reduces the computational cost required for model updates while enabling rapid adaptation to dynamic changes in the power system, thus maintaining the accuracy and timeliness of predictions over the long term.

[0080] Specifically, when the deviation exceeds a preset threshold at multiple consecutive moments (e.g., three consecutive 15-minute intervals), it indicates that the model's real-time dynamic capture capability may be lagging. At this time, the system automatically uses a sliding window of data consisting of real-time power load data from recent consecutive time periods (e.g., the past 24 hours) and corresponding real-time external factor data to fine-tune the parameters of the second neural network model that generates real-time dynamic feature vectors in the prediction model.

[0081] For example, if a sudden thunderstorm occurs on a summer afternoon, causing a sharp drop in temperature and a subsequent surge in power load, and the predicted values ​​for several consecutive moments are lower than the actual measured values ​​with deviations exceeding a set threshold, the system will automatically trigger an update mechanism. The system utilizes a sliding window of data from several hours before and after the thunderstorm to rapidly incrementally train only the second neural network model. This allows the model to quickly learn the new operating pattern of a sudden drop in temperature accompanied by a rebound in load, while keeping the parameters of the first neural network model, responsible for extracting historical long-term cycle patterns, unchanged. This ensures timely improvement in predictive capability while maintaining the overall stability of the model.

[0082] It should be understood that the ultra-short-term power load forecasting method in this embodiment of the invention uses a forecasting model. Figure 5 This is a schematic structural block diagram of a prediction model according to an embodiment of the present invention, such as... Figure 5 As shown, the prediction model 20 may include three parts: a historical pattern extraction module 21, a real-time dynamic capture module 22, and a feature fusion and prediction generation module 23. Wherein:

[0083] The historical pattern extraction module 21 is constructed based on the first neural network model 211 and is used to generate historical pattern feature vectors that reflect typical daily load patterns and trends. These vectors represent the steady-state experience of power system operation.

[0084] The real-time dynamic capture module 22, built on the second neural network model 221, is used to generate a real-time dynamic feature vector that captures the latest changes of the day and reflects real-time deviations and external influences. This vector represents the transient state of the current operation.

[0085] The feature fusion and prediction generation module 23 has a prediction output layer 231. By introducing a dynamic fusion mechanism based on the degree of difference, it adaptively adjusts the fusion weights of the two types of features according to the degree of difference between the real-time dynamic feature vector and the historical pattern feature vector, and fuses the real-time dynamic feature vector and the historical pattern feature vector according to the fusion weights. Finally, the prediction output layer 231 outputs the ultra-short-term power load prediction results for the remaining period of the day.

[0086] This embodiment also provides a computer program product 11, a computer-readable storage medium 12, and a computer device 10. Figure 6This is a schematic diagram of a computer program product according to an embodiment of the present invention. Figure 7 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention. Figure 8 This is a schematic block diagram of a computer device according to an embodiment of the present invention.

[0087] Computer program product 11 includes computer program 111, which, when executed by processor 101, implements any of the aforementioned ultra-short-term power load forecasting methods. Computer-readable storage medium 12 stores the aforementioned computer program 111, which, when executed by processor 101, implements any of the aforementioned ultra-short-term power load forecasting methods. Computer device 10 may include memory 102, processor 101, and computer program 111 stored in memory 102 and running on processor 101.

[0088] The computer program 111 used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, status setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​and procedural programming languages.

[0089] Computer program 111 may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, to perform aspects of the invention, electronic circuits including, for example, programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) may execute computer-readable program instructions to personalize the electronic circuits by utilizing state information of computer-readable program instructions.

[0090] For the purposes of this embodiment, computer program product 11 is a related product containing computer program 111. Computer-readable storage medium 12 is a tangible device capable of holding and storing computer program 111, and can be any device capable of containing, storing, communicating, propagating, or transmitting computer program 111 for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage medium 12 include: portable computer disks, hard disks, random access memory 102 (RAM), read-only memory 102 (ROM), erasable programmable read-only memory 102 (EPROM or flash memory), static random access memory 102 (SRAM), portable optical disc read-only memory 102 (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, mechanical encoding device, and any suitable combination thereof.

[0091] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A method for ultra-short-term power load forecasting, characterized in that, include: Acquire historical power load data and historical external factor data stored in the historical database, and generate historical pattern feature vectors that characterize typical daily load patterns; Acquire real-time power load data and real-time external factor data from the start of the day to the present, and generate a real-time dynamic feature vector characterizing the trend of power load changes on the day; Dynamically determine the fusion weights used to fuse the historical pattern feature vector and the real-time dynamic feature vector; The historical pattern feature vector and the real-time dynamic feature vector are weighted and fused using the fusion weights to obtain a fused feature vector. Based on the fused feature vector, the ultra-short-term power load forecast results from the current time to the end of the day are output; The step of generating historical pattern feature vectors representing typical daily load patterns includes: inputting the historical power load data and the historical external factor data into a first neural network model; The steps for generating a real-time dynamic feature vector representing the daily electricity load change trend include: inputting the real-time electricity load data and the real-time external factor data into a second neural network model; After the step of outputting the ultra-short-term power load forecast result from the current time to the end of the day based on the fused feature vector, the method further includes: In response to receiving new real-time load data, the deviation between the real-time load value of the new real-time load data and the load forecast value corresponding to the current time in the previous forecast is calculated; When the deviation exceeds a preset threshold at multiple consecutive times, the parameters of the second neural network model that generates the real-time dynamic feature vector in the prediction model are adjusted using sliding window data composed of real-time power load data and corresponding real-time external factor data from recent consecutive periods.

2. The ultra-short-term power load forecasting method according to claim 1, characterized in that, The steps for generating historical pattern feature vectors representing typical daily load patterns also include: The first neural network model is used to analyze the periodic variation pattern of the historical power load data at different time scales, and the periodic variation pattern at each time scale is corrected according to the historical external factor data, thereby generating the historical pattern feature vector.

3. The ultra-short-term power load forecasting method according to claim 1, characterized in that, The steps for generating a real-time dynamic feature vector representing the daily electricity load change trend also include: The second neural network model is used to extract the temporal evolution characteristics of the real-time power load data and the dynamic response characteristics of the real-time external factor data; The continuous trend of daily load change is captured based on the time-series evolution characteristics, and the real-time impact of external factor fluctuations on the load is quantified based on the dynamic response characteristics to identify the deviation between the real-time power load data and the typical daily load pattern, thereby generating the real-time dynamic feature vector.

4. The ultra-short-term power load forecasting method according to claim 1, characterized in that, The step of dynamically determining the fusion weights for fusing the historical pattern feature vector and the real-time dynamic feature vector includes: Calculate the difference between the real-time dynamic feature vector and the historical pattern feature vector; The first weight of the historical pattern feature vector and the second weight of the real-time dynamic feature vector are generated based on the difference degree, wherein the greater the difference degree, the higher the proportion of the second weight relative to the first weight.

5. The ultra-short-term power load forecasting method according to claim 4, characterized in that, The steps of generating the first weight of the historical pattern feature vector and the second weight of the real-time dynamic feature vector based on the difference include: The first weight and the second weight are calculated based on the difference using an attention mechanism; or After concatenating the historical pattern feature vector with the real-time dynamic feature vector, the first weight and the second weight are obtained by mapping them through a preset neural network based on the degree of difference.

6. The ultra-short-term power load forecasting method according to claim 1, characterized in that, After the step of outputting the ultra-short-term power load forecast result from the current time to the end of the day based on the fused feature vector, the method further includes: The ultra-short-term power load forecast results, the forecast time corresponding to the ultra-short-term power load forecast results, and real-time external factor data are associated and stored in the historical database.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the ultra-short-term power load forecasting method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the ultra-short-term power load forecasting method according to any one of claims 1 to 6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the ultra-short-term power load forecasting method according to any one of claims 1 to 6.

Images