Charging pile load prediction method based on lstm and multi-source feature fusion

Abstract

The application provides a charging pile load prediction method based on LSTM and multi-source feature fusion, relates to the technical field of data prediction, and comprises the following steps: collecting multi-source indication data related to the load of a target charging pile, performing feature fusion on the multi-source indication data, and forming a sample data sequence; performing multi-scale division on the sample data sequence by using multiple time length analysis windows, obtaining multi-scale sample data sequences, and respectively constructing and training multiple long short-term memory network prediction channels corresponding to the analysis windows according to the multi-scale sample data sequences to obtain a multi-channel prediction component; updating real-time multi-source indication data of the target charging pile, and performing joint load prediction on the real-time multi-source indication data and the multi-channel prediction component to obtain predicted load information. The technical problem of low accuracy of charging pile load prediction in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of data prediction, and in particular to a method for predicting the load of charging piles based on LSTM and multi-source feature fusion. Background Technology

[0002] With the increasing popularity of electric vehicles, charging stations, as the core infrastructure for electric vehicle energy replenishment, have been widely deployed in urban business districts, residential communities, transportation hubs, and other scenarios. Accurately predicting the load on charging stations is not only a key basis for supporting the power grid to achieve coordinated scheduling of power generation, grid, load, and storage, but also an important prerequisite for optimizing the operation and management of charging stations and improving the user charging experience.

[0003] However, existing charging pile load forecasting methods mostly rely on single historical load data for forecasting, without effectively integrating key influencing factors such as time period and external environment. This results in a lack of feature dimensions, making it difficult to characterize the complex driving mechanism of load changes. Consequently, the accuracy of charging pile load forecasting is not high, making it difficult to meet the precision requirements of power grid dispatching and charging pile operation.

[0004] Therefore, there is an urgent need for a charging pile load prediction method based on LSTM and multi-source feature fusion that can integrate key features from multiple sources and take into account real-time performance, so as to break through the existing technical bottlenecks and provide reliable support for the efficient operation of charging piles and the safe and stable operation of the power grid. Summary of the Invention

[0005] This invention addresses the technical problem of low accuracy in charging pile load prediction in existing technologies by providing a charging pile load prediction method based on LSTM and multi-source feature fusion.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a charging pile load prediction method based on LSTM and multi-source feature fusion, including:

[0008] Collect multi-source feature data related to the load of the target charging pile, and perform feature fusion on the multi-source feature data to form sample time series data;

[0009] The sample time series data is divided into multiple scales using analysis windows of various time lengths to obtain multi-scale sample time series data. Based on the multi-scale sample time series data, multiple long short-term memory network prediction channels corresponding to the analysis windows are constructed and trained to obtain multi-channel prediction components.

[0010] Update the real-time multi-source feature data of the target charging pile, and combine the real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction and obtain predicted load information.

[0011] The beneficial effects of this invention are:

[0012] Compared to existing technologies, this application first collects multi-source feature data related to the load of the target charging pile, and then fuses these multi-source feature data to form sample time-series data. This breaks through the information limitations of a single data source, transforming scattered influencing factors into structured, time-series input samples, providing multi-source and reliable structured data for subsequent multi-scale analysis and model prediction. Secondly, it uses analysis windows of various time lengths to divide the sample time-series data into multiple scales, obtaining multi-scale sample time-series data. Based on this multi-scale sample time-series data, it constructs and trains multiple long short-term memory network prediction channels corresponding to the analysis windows, obtaining multi-channel prediction components. This avoids the attention diversion problem of a single model in multi-scale feature learning, providing multi-dimensional model support for subsequent joint load prediction combined with real-time data. Finally, it updates the collected real-time multi-source feature data of the target charging pile and combines this real-time multi-source feature data with the multi-channel prediction components for joint load prediction, obtaining predicted load information. This enables the model to adapt to real-time dynamics, integrates the prediction advantages of different time scales, and ensures that the final output predicted load information is both reliable and accurate, and closely reflects changes in the actual scenario.

[0013] Through the above technical solution, this application overcomes the shortcomings of traditional methods that rely on single data and lack feature dimensions by collecting multi-source feature data related to the target charging pile load. It can comprehensively capture multiple load driving factors, providing complete data support for prediction. By employing analysis windows of various time lengths to divide multi-scale sample time-series data and constructing and training corresponding long short-term memory network prediction channels to form a multi-channel prediction component, it overcomes the problem that single-time-scale models cannot adequately account for changes in trends over different periods. Furthermore, by updating and collecting real-time multi-source feature data and combining it with the multi-channel prediction component to conduct joint load prediction, it solves the problems of insufficient real-time performance and large deviations in handling dynamic scenarios inherent in traditional methods. In this way, it ensures that the charging pile load prediction results can adapt to dynamic factors such as changes in user behavior and sudden environmental impacts in real time, improving the accuracy and real-time performance of the prediction results. This provides accurate and reliable load data for grid-source-load-storage coordinated scheduling, charging pile resource allocation optimization, and improvement of user charging experience. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the charging pile load prediction method based on LSTM and multi-source feature fusion provided by the present invention.

[0015] Figure 2 This is a schematic diagram of the process for obtaining predicted load information in the charging pile load prediction method based on LSTM and multi-source feature fusion provided by the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0019] Examples, such as Figure 1 As shown, this embodiment of the invention provides a charging pile load prediction method based on LSTM and multi-source feature fusion, including:

[0020] S10: Collect multi-source feature data related to the load of the target charging pile, and perform feature fusion on the multi-source feature data to form sample time series data.

[0021] The load on charging piles is affected by a combination of factors, including historical operating status, time cycle patterns, user charging behavior, and changes in the external environment. A single type of data is insufficient to fully characterize its complex fluctuations, and the effectiveness of subsequent multi-scale LSTM prediction models highly depends on the completeness and temporal correlation of the input data.

[0022] To address the aforementioned issues, this application collects multi-source feature data related to the load of the target charging pile and performs feature fusion on the multi-source feature data to form sample time-series data.

[0023] Specifically, step S10 in the method includes:

[0024] Collect multi-source recorded information of the target charging pile, including historical load sequence information, time period information, user behavior information, and external environment information;

[0025] Based on feature engineering and the multi-source record information, the multi-source feature data is extracted and concatenated to obtain the sample time series data.

[0026] In this embodiment of the application, multi-source recording information of the target charging pile is first collected to cover various key factors affecting the load change of the charging pile. The multi-source recording information includes at least historical load sequence information, time period information, user behavior information and external environment information.

[0027] Historical load sequence information refers to the actual load data of the target charging pile during its past operation, such as charging power records every 15 minutes, daily total charging volume statistics, and the specific time and duration of load peaks. Historical load sequence information directly reflects the historical fluctuation pattern of the target charging pile's load and is an important basis for the model to learn load change trends.

[0028] The time period information refers to the periodic characteristics of the target charging pile's load, including basic time attributes such as hour, date, and day of the week, and period labels such as weekday / weekend, holiday / ordinary day, season, and morning / evening peak hours. Since charging pile loads often exhibit clear periodic patterns—for example, peak loads before 8 am and after 7 pm on weekdays due to commuting demand, and increased loads on weekend afternoons due to leisure travel demand—time period information helps the model accurately identify load change patterns under different periods.

[0029] The user behavior information includes user charging frequency, preferred charging time (e.g., office workers often choose to charge at night or during lunch break), user type distribution (e.g., the charging ratio of private cars, ride-hailing vehicles, and commercial vehicles), and charging duration preference (the ratio of fast charging to slow charging). User behavior information directly determines the timing and intensity of charging demand, and the model can use this information to capture load fluctuations caused by human factors.

[0030] External environmental information refers to information about the environment in which the target charging station is located, including meteorological data (such as temperature, humidity, wind speed, etc.), surrounding environment (such as whether the location is a residential area / commercial area / office building, etc.), and regional factors (such as whether there are charging subsidies). External environmental information indirectly affects the charging station load by influencing users' travel plans and charging decisions, and is key data for improving the robustness of model predictions.

[0031] Secondly, since the collected multi-source record information is mostly raw, unstructured, or multi-level data, it cannot be directly input into the LSTM model. Therefore, the multi-source record information is transformed into standardized, structured multi-source feature data through feature engineering of the system, and then spliced ​​together to form sample time series data.

[0032] For example, the multi-source record information is first cleaned. Missing values ​​(such as missing power records for a certain period) can be filled using methods like mean imputation or linear interpolation. Outliers (such as extremely high or zero power records far exceeding the normal range) can be identified and removed using methods like box plots or the 3σ principle to ensure data accuracy. Next, feature transformation is performed, converting non-numerical information into numerical features recognizable by the model. For instance, one-hot encoding can convert non-numerical information into quantifiable numerical features. Then, Pearson correlation analysis and mutual information entropy methods are used to filter features highly correlated with load values, removing irrelevant or redundant features. Features of different magnitudes are standardized to the same scale using Min-Max or Z-Score standardization to avoid the impact of data magnitude differences on the model's recognition performance. In this way, standard multi-source feature data is obtained.

[0033] For example, the multi-source feature data after feature engineering needs to be concatenated in chronological order to form sample time-series data suitable for LSTM model input. For instance, a reasonable time interval (such as 15 minutes, 30 minutes, or 1 hour) is chosen as the time-series unit. Each time unit corresponds to a set of all multi-source feature data at that time point. If a 15-minute interval is used, a certain time point (such as 18:00-18:15) includes: historical load sequence information after feature engineering, time period information, user behavior information, and external environment information. Thus, according to the time-series information, the multi-source feature data of each time interval is concatenated to obtain sample time-series data. This sample time-series data is equivalent to a multi-dimensional feature matrix arranged in order along the time axis, with rows representing different time points and columns representing various types of multi-source feature data.

[0034] In summary, compared to existing technologies, this application collects multi-source feature data related to the load of the target charging pile and performs feature fusion on the multi-source feature data to form sample time-series data. This breaks through the information limitations of a single data source, transforming dispersed influencing factors into structured, time-series input samples, providing multi-source, reliable structured data for subsequent multi-scale analysis and model prediction.

[0035] S20: The sample time series data is divided into multiple scales using analysis windows of various time lengths to obtain multi-scale sample time series data. Based on the multi-scale sample time series data, multiple long short-term memory network prediction channels corresponding to the analysis windows are constructed and trained to obtain multi-channel prediction components.

[0036] The load on charging piles exhibits differentiated characteristics across different time dimensions. In the short term, it may fluctuate dramatically due to immediate user demand; in the medium term, it shows a daily / weekly cyclical pattern; and in the long term, it exhibits trend changes due to factors such as seasons. Traditional single-time-scale models are difficult to take into account multi-dimensional patterns and are prone to insufficient learning of features at certain scales.

[0037] To address the aforementioned issues, this application employs analysis windows of varying time lengths to perform multi-scale segmentation of the sample time-series data, thereby obtaining multi-scale sample time-series data. Based on the multi-scale sample time-series data, multiple long short-term memory network prediction channels corresponding to the analysis windows are constructed and trained to obtain multi-channel prediction components.

[0038] Specifically, step S20 in the method includes:

[0039] Based on the historical load prediction records of the target charging piles, the time lengths of multiple historical analysis windows are extracted, and analysis windows of various time lengths are defined accordingly.

[0040] Based on the defined analysis windows of various time lengths, the sample time series data are subjected to sliding segmentation processing to generate the corresponding multi-scale sample time series data, wherein the multi-scale sample time series data includes multiple sample time series data subsets.

[0041] In this embodiment, firstly, based on the historical load forecast records of the target charging pile, data mining is used to extract meaningful time-dimensional patterns. Then, the time lengths of multiple historical analysis windows are determined, and corresponding analysis windows of various time lengths are defined. For example, by mining the historical load forecast records of the target charging pile, it may be found that: within a time span of approximately one hour, load fluctuations are mainly affected by users' temporary charging needs; within a time span of one day, load changes highly coincide with commuting peak hours and commercial activity periods; and within a time span of one week, load differences are clearly reflected in the differences in user behavior between weekdays and weekends. Based on these time-dimensional patterns extracted from the historical load forecast records, the time lengths of multiple historical analysis windows are determined, such as one hour, one day, and one week, thus ensuring that each analysis window accurately corresponds to a type of time-dimensional pattern.

[0042] Secondly, based on the defined analysis windows of various time lengths, sliding segmentation is performed on the sample time series data to generate corresponding multi-scale sample time series data. The multi-scale sample time series data includes multiple subsets of sample time series data. Sliding segmentation refers to dividing the continuous sample time series data into multiple non-overlapping or partially overlapping subsequences, starting from the beginning of the sample time series data and sliding forward at preset step sizes (usually 1 / 2 or 1 / 4 of the window time length) using the selected analysis window time length as a fixed unit.

[0043] For example, if the sample time series data is spliced ​​together with 15-minute intervals, when using an analysis window of 1 hour (containing 4 data points) and a step size of 30 minutes (containing 2 data points), the first sliding segment will cut off the 1st to 4th data points as the first subsequence, and after the second sliding, the 3rd to 6th data points will be cut off as the second subsequence, and so on, eventually generating a set of subsequences covering the entire time axis with a 1-hour analysis window. Similarly, by continuing to use other selected analysis window time lengths as fixed units, several subsequence sets of different time lengths can be cut out. These subsequence sets of different time lengths together constitute multi-scale sample time series data. Each subsequence set focuses on the load characteristics of a specific time length, which can provide targeted data for subsequent multi-scale model training.

[0044] Furthermore, the step of "constructing and training multiple long short-term memory network prediction channels corresponding to the analysis window based on the multi-scale sample time-series data, and obtaining a multi-channel prediction component" includes:

[0045] Based on analysis windows of various time lengths, the same number of prediction channels for the Long Short-Term Memory network are constructed accordingly.

[0046] Using multiple subsets of the sample time-series data as training data, multiple prediction channels of the Long Short-Term Memory Network are trained respectively, and the prediction residual of each prediction channel of the Long Short-Term Memory Network is calculated accordingly.

[0047] When each of the predicted residuals is less than the preset individual accuracy threshold, the multiple prediction channels of the long short-term memory network are determined to have passed the pre-training verification and are integrated and output as the multi-channel prediction component.

[0048] In this embodiment, firstly, based on analysis windows of various time lengths, the same number of Long Short-Term Memory (LSTM) network prediction channels are constructed accordingly. That is, a separate LTM network prediction channel is designed and built for each analysis window of different time lengths. For example, the network structure of each LTM network prediction channel can be adapted to the complexity of the load characteristics for the corresponding time length. For instance, for an LTM network prediction channel with a 1-hour analysis window, a shallower network layer and fewer hidden neurons can be designed to quickly capture high-frequency fluctuations in short-term load. For an LTM network prediction channel with a 1-week analysis window, the network depth and number of neurons can be increased to enhance the ability to learn long-term load trends. Thus, through differentiated model architectures, the problem of attentional distraction that occurs when a single LTM network model simultaneously learns short-, medium-, and long-term multi-scale features can be avoided, ensuring that each LTM network prediction channel can focus on the load patterns of its corresponding time length.

[0049] Secondly, multiple long short-term memory (LSM) network prediction channels are trained using multiple subsets of sample time-series data as training data, and the prediction residuals of each LSM network prediction channel are calculated accordingly. For example, the sample time-series data subsets are matched with LSM network prediction channels corresponding to the same time length, and each is trained independently. For instance, a 1-hour sample time-series data subset is used to train an LSM network prediction channel for a 1-hour analysis window, a 1-day sample time-series data subset is used to train an LSM network prediction channel for a 1-day analysis window, a 1-week sample time-series data subset is used to train an LSM network prediction channel for a 1-week analysis window, and so on.

[0050] During training, the sample time-series data subset can be divided into training set, validation set, and test set in a ratio of 7:1.5:1.5. The training set uses the backpropagation algorithm to iteratively optimize the network weights, biases, and other parameters of each long short-term memory network prediction channel. The validation set dynamically adjusts hyperparameters such as learning rate and number of iterations by evaluating the prediction effect to avoid model overfitting. After training, the prediction residuals of each long short-term memory network prediction channel are validated. The sample time-series data of the test set is input into the long short-term memory network prediction channel, and the mean absolute error (MAE) and root mean square error (RMSE) of the predicted loading value output by the long short-term memory network prediction channel and the actual loading value in the validation set are calculated as the prediction residuals.

[0051] Finally, when each prediction residual is less than a preset individual precision threshold, the multiple prediction channels of the Long Short-Term Memory network are determined to have passed the pre-training verification and are integrated into a multi-channel prediction component. The individual precision threshold can be dynamically set according to the actual scenario and computing power accuracy; for example, the mean absolute error (MAE) can be set to ≤3%.

[0052] For example, each Long Short-Term Memory (LSTM) prediction channel can only be considered to have passed the pre-training verification when the prediction residuals of all LTM prediction channels are less than the preset individual accuracy threshold (e.g., MAE ≤ 3%). This ensures that each LTM prediction channel has sufficient prediction accuracy and avoids the overall effect of subsequent multi-channel joint prediction being affected by excessive single-channel errors. Then, the several LTM prediction channels that have passed the pre-training verification are integrated and output as a multi-channel prediction component. The multi-channel prediction component integrates the independent prediction capabilities and advantages of each LTM prediction channel, possesses accurate prediction capabilities for different time lengths, and has multi-dimensional synergistic and complementary advantages.

[0053] It should be noted that Long Short-Term Memory (LSTM) networks are well-known existing technologies in this field, and their core principles, such as gating mechanisms and temporal information transmission logic, have been fully disclosed in relevant technical literature and practice. Regarding the specific architecture design and technical implementation details of the LSTM network prediction channel in this method, those skilled in the art can flexibly adjust them based on the well-known LSTM technology framework, combined with the load data characteristics and prediction accuracy requirements of the target charging pile; therefore, no further detailed explanation is necessary here.

[0054] In summary, compared to existing technologies, this application employs analysis windows of multiple time lengths to divide the sample time-series data into multiple scales, obtaining multi-scale sample time-series data. Based on this multi-scale sample time-series data, multiple long short-term memory network prediction channels corresponding to the analysis windows are constructed and trained to obtain multi-channel prediction components. This avoids the attentional distraction problem of a single model in multi-scale feature learning, providing multi-dimensional model support for subsequent joint load prediction using real-time data.

[0055] S30: Update the real-time multi-source feature data of the target charging pile, and combine the real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction and obtain predicted load information.

[0056] The aforementioned steps train a multi-channel prediction component, which can combine real-time user behavior and environmental changes of the target charging station to predict load information.

[0057] To address the aforementioned issues, this application updates the real-time multi-source feature data of the target charging piles and combines the real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction and obtain predicted load information.

[0058] Specifically, such as Figure 2 As shown, step S30 in the method includes:

[0059] Using the longest analysis window as the acquisition constraint, the real-time multi-source recording information of the target charging pile is updated and the real-time multi-source feature data is extracted based on feature engineering.

[0060] The analysis window with multiple time lengths is used to perform multi-scale partitioning of the real-time multi-source feature data to obtain a multi-scale prediction input set;

[0061] The multi-channel prediction component is subjected to post-consistency verification based on the multi-scale prediction input set. If the verification passes, joint load fusion is performed based on the post-consistency verification result to obtain the predicted load information.

[0062] In this embodiment, the longest analysis window is first used as the acquisition constraint to update the real-time multi-source recording information of the target charging pile, and real-time multi-source feature data is extracted based on feature engineering. For example, if the longest time length of the analysis window is one week, real-time multi-source recording information including historical load sequence information, time period information, user behavior information, and external environment information of the most recent week is collected to ensure that the input requirements of all time lengths are covered. After the same feature engineering processing as in step S10, real-time multi-source feature data with the same sample time series data structure is formed, ensuring consistency with the training sample structure of the long short-term memory network prediction channel in the multi-channel prediction component.

[0063] Secondly, the real-time multi-source feature data is divided into multiple scales based on analysis windows of various time lengths to obtain a multi-scale prediction input set. For example, following the same sliding segmentation processing method as step S20, the real-time multi-source feature data is divided into multi-scale prediction input sets corresponding one-to-one with the prediction channels of the Long Short-Term Memory network. For instance, if the real-time multi-source feature data is spliced ​​together with 15-minute time intervals, using an analysis window of 1 hour (containing 4 data points) and a step size of 30 minutes (containing 2 data points), the first sliding segmentation will extract data points 1-4 as the first prediction input set, the second sliding segmentation will extract data points 3-6 as the second prediction input set, and so on, ultimately generating a 1-hour analysis window subsequence set covering the entire time axis. The same method is then used with analysis windows of 1 day and 1 week, with corresponding step sizes, to perform sliding segmentation and obtain the multi-scale prediction input set.

[0064] Finally, post-consistency verification is performed on the multi-channel prediction components based on the multi-scale prediction input set. If the verification passes, joint load fusion is performed based on the post-consistency verification results to obtain the predicted load information. Post-consistency verification verifies whether the prediction results of different Long Short-Term Memory (LSTM) network prediction channels are contradictory. If the verification passes, the multi-channel results are integrated through joint load fusion to ultimately generate the predicted load information. This ensures that real-time prediction utilizes the specialized capabilities of multiple channels while mitigating the limitations of a single channel through verification and fusion.

[0065] Specifically, the "post-consistency verification of the multi-channel prediction component based on the multi-scale prediction input set" includes:

[0066] Based on the correspondence between the analysis window and the multiple long short-term memory network prediction channels, the multi-scale prediction input set is input into the multiple long short-term memory network prediction channels respectively and iterative prediction is performed for a preset number of consecutive times to obtain the multi-scale prediction output set.

[0067] Traverse each subset of the multi-scale prediction output set and perform curve fitting to obtain the corresponding set of typical output curves.

[0068] The typical output curves are randomly combined, and the mutual residuals between the typical output curves are calculated based on the combination results.

[0069] If each of the aforementioned mutual residuals meets the preset consistency threshold, then the post-consistency verification is passed, and joint load fusion is allowed.

[0070] In this embodiment, firstly, based on the correspondence between the analysis window and multiple long short-term memory network prediction channels, the multi-scale prediction input set is input into each of the multiple long short-term memory network prediction channels, and iterative prediction is performed for a preset number of consecutive iterations to obtain the multi-scale prediction output set. The preset number of iterations is performed to reduce random errors in the long short-term memory network prediction channels and improve the accuracy and stability of the prediction results. The preset number of iterations can be dynamically set according to the actual application scenario and prediction accuracy requirements.

[0071] For example, the analysis window of the multi-scale prediction input set is matched with the corresponding long short-term memory network prediction channel. The multi-scale prediction input set is input into the corresponding multiple long short-term memory network prediction channels for iterative (e.g., 3 rounds) prediction. Each long short-term memory network prediction channel outputs a prediction output subset containing multiple prediction results, which are then summarized to form the multi-scale prediction output set.

[0072] Secondly, curve fitting is performed on each subset of predicted outputs in the multi-scale predicted output set to obtain the corresponding set of typical output curves. For example, the mean curve, median curve, or polynomial can be used to fit curves to each subset of predicted outputs to obtain several typical output curves, which are then summarized to form a set of typical output curves. Each typical output curve corresponds to the core prediction trend of a prediction channel of a long short-term memory network, eliminating noise interference from single predictions.

[0073] Next, a random combination is performed based on the set of typical output curves, and the mutual residuals between multiple typical output curves are calculated based on the combination results. For example, the typical output curves fitted by the prediction channels of the three long short-term memory networks are randomly combined pairwise, and the prediction deviation of the corresponding time points of the two typical output curves in the combination is calculated as the mutual residuals between the typical output curves.

[0074] Finally, if each mutual residual meets the preset consistency threshold, the post-consistency verification is deemed to have passed and joint load fusion is allowed. Otherwise, if any mutual residual does not meet the preset consistency threshold, the post-consistency verification is deemed to have failed. It is necessary to backtrack and investigate issues such as the quality of real-time multi-source recorded information acquisition, the rationality of multi-scale partitioning, and the status of long short-term memory network prediction channels. The post-consistency verification will be re-executed after the issues are resolved.

[0075] The preset consistency threshold can be dynamically set based on factors such as the historical load fluctuation range of the target charging pile and the accuracy requirements of the prediction task. For example, the prediction deviation can be set to ≤8% as the preset consistency threshold. Those skilled in the art can flexibly set it according to specific application requirements.

[0076] In this way, by performing post-conformity verification, the reliability of the prediction results from multiple long short-term memory network prediction channels is ensured, thereby improving the accuracy and reliability of the final predicted load information.

[0077] Furthermore, the step of "performing joint load fusion based on post-consistency verification results to obtain the predicted load information" includes:

[0078] Based on the results of the post-consistency verification, the typical output curve set is extracted;

[0079] The typical output curve set is used for segmented weighted fusion, wherein the weight of the typical output curve is monotonically positively correlated with the time length of the corresponding analysis window;

[0080] The output segmented weighted fusion result is the predicted load information.

[0081] In this embodiment of the application, a set of typical output curves that have passed the post-consistency verification is first extracted based on the post-consistency verification results.

[0082] Secondly, segmented weighted fusion is performed based on the typical output curve set. The weight of the typical output curve is monotonically positively correlated with the time length of the corresponding analysis window. That is, the longer the time length of the analysis window, the higher the weight of the typical output curve. This is because the longer the time length of the analysis window, the better it reflects the long-term trend of the target charging pile load. The shorter the time length of the analysis window, the lower the weight of the typical output curve, in order to supplement the details of short-term fluctuations.

[0083] Finally, the segmented weighted fusion result is output as the predicted load information. Specifically, since the segmented weighted fusion results from different time periods may experience numerical jumps or abrupt trend changes at their segmentation points during splicing due to the switching of weights on curves at different scales, it is necessary to splice the segmented weighted fusion results based on the time series relationship, identify the splicing breaks, and use a preset smoothing function to smooth multiple splicing breaks, making the final curve continuous and smooth. The smoothed result is defined as the predicted load information. The preset smoothing function can be dynamically selected according to the actual situation; for example, a moving average suitable for short-term detail smoothing, exponential smoothing suitable for trend continuity, or polynomial fitting suitable for complex curve transitions can be selected.

[0084] For example, the weighted fusion results of each segment are spliced ​​together in chronological order. The splicing breakpoints are identified by detecting features such as the numerical difference and changes in the first derivative at the junctions of adjacent segments. For the identified breakpoints, a preset smoothing function is used for smoothing. For instance, if the moving average method is selected, the predicted values ​​at 3-5 time points before and after the breakpoint can be weighted and averaged. For example, the weight at the central breakpoint is set to 0.3, decreasing by 0.1 towards both sides, thus eliminating abrupt changes through a gradual transition. If exponential smoothing is selected, the smoothing coefficient can be set to 0.6-0.8, correcting the first 3-4 points of the subsequent segment to control the deviation from the endpoint value of the previous segment within 5%, ultimately making the overall curve continuously smooth. The smoothed result is defined as the predicted load information. In this way, the core impact scale of different time periods is preserved while ensuring the consistency of the results.

[0085] Specifically, the statement that "the weighted average of the typical output curves is monotonically positively correlated with the time length of the corresponding analysis window" includes:

[0086] Using the end point of each analysis window as the dividing point, the typical output curve set is segmented to obtain multiple segmented curve subsets, wherein the number of segmented output curves included in each segmented subset decreases along the time direction.

[0087] Traverse multiple piecewise curve subsets, normalize the preset window weights corresponding to multiple analysis windows in each piecewise curve subset, and obtain multiple piecewise weighted weight sets.

[0088] Based on multiple segmented weighted weight sets, multiple segmented curve subsets are weighted and fused respectively to obtain the segmented weighted fusion result.

[0089] In this embodiment, the typical output curve set is first segmented using the end point of each analysis window as the dividing point to obtain multiple segmented curve subsets. The number of segmented output curves in each segmented subset decreases along the time direction. For example, using the end point of each analysis window as the time dividing point, if the analysis window length is 1 hour, 1 day, or 1 week, the corresponding end points are current time + 1 hour, current time + 1 day, and current time + 1 week, respectively. These three end points divide the future time axis into three continuous intervals: 0~1 hour, 1 hour~1 day, and 1 day~1 week, thus obtaining multiple segmented curve subsets.

[0090] For example, due to the differences in the time coverage capabilities of different analysis windows, the number of typical output curves included in each interval decreases along the time direction: In the 0-1 hour interval, the typical output curves of the 1 hour, 1 day, and 1 week analysis windows can completely cover this interval, so the segmented curve subset contains 3 typical output curves; In the 1 hour to 1 day interval, the typical output curves of the 1 hour window only cover 0-1 hour and cannot extend to this interval, so the subset only retains 2 typical output curves of the 1 day and 1 week analysis windows; In the 1 day to 1 week interval, the typical output curves of the 1 hour and 1 day analysis windows cannot be covered, and only 1 typical output curve is included. In this way, by decreasing the number, the interference of invalid typical output curves on the fusion results is eliminated, ensuring that the analysis of each interval is based on the adapted scale features.

[0091] Secondly, the process iterates through multiple piecewise curve subsets, normalizes the preset window weights corresponding to multiple analysis windows in each piecewise curve subset, and obtains multiple piecewise weighted weight sets. For example, the window weights are first preset based on the time length of the analysis window. Generally, the longer the time length of the analysis window, the more stable the characterization of the long-term trend, and the higher the preset window weight. For example, preset window weights of 0.2, 0.3, and 0.5 are preset for windows with time lengths of 1 hour, 1 day, and 1 week, respectively.

[0092] For example, subsequently, for each piecewise curve subset, the preset window weights corresponding to the typical output curves in that piecewise curve subset are extracted and normalized. For instance, for the three typical output curves in the 0-1 hour interval, the sum of the preset window weights is 0.2 + 0.3 + 0.5 = 1, and no adjustment is needed; (0.2, 0.3, 0.5) is directly used as the weighted weight set for this interval. For the two typical output curves (1 day and 1 week) in the 1 hour to 1 day interval, the sum of the preset window weights is 0.3 + 0.5 = 0.8, and after normalization, the weights are 0.3 / 0.8 = 0.375 and 0.5 / 0.8 = 0.625, respectively. (0.375, 0.625) is used as the weighted weight set for this interval. For the one typical output curve (1 week) in the 1 day to 1 week interval, the preset window weight is 0.5, and after normalization, it is 1.0. 1.0 is used as the weighted weight set for this interval. In this way, normalization ensures that the weight allocation within each interval retains the scale priority while generating a reasonable fusion value through weighted summation.

[0093] Finally, based on multiple segmented weighted weight sets, multiple segmented curve subsets are weighted and fused separately to obtain the segmented weighted fusion result. Specifically, for each segmented curve subset, the predicted value of each time point of the curve needs to be weighted and calculated in conjunction with the corresponding segmented weighted weight set.

[0094] For example, using the segmented weighted weight set from the previous step, within the 0-1 hour interval, the fusion value at any time point t = (predicted value of typical output curve t at 1 hour × 0.2) + (predicted value of typical output curve t at 1 day × 0.3) + (predicted value of typical output curve t at 1 week × 0.5). This retains the detailed capture of short-term load fluctuations by the 1-hour analysis window while avoiding short-term noise interference through the weight constraints of the 1-day and 1-week analysis windows. Within the 1-hour to 1-day interval, the fusion value at any time point t = (predicted value of typical output curve t at 1 day × 0.375) + (predicted value of typical output curve t at 1 week × 0.5). The weighted average value of the line at point t is multiplied by 0.625. By increasing the weight of the long window, the influence of the daily cycle pattern on the load is strengthened. Within the range of 1 day to 1 week, the fusion value of any time point t is equal to the predicted value of the typical output curve at point t in 1 week multiplied by 1.0. This relies entirely on the trend prediction of the long window to ensure the robustness of the long-term results. Finally, the calculation results of the three intervals are spliced ​​together in chronological order to obtain the complete segmented weighted fusion result. The segmented weighted fusion result covers the load characteristics of the short, medium and long term, and achieves a reasonable balance of information at different scales through weight allocation, providing a coherent basic data for subsequent smoothing processing.

[0095] Furthermore, the phrase "performing joint load fusion based on post-consistency verification results to obtain the predicted load information" also includes:

[0096] Based on the post-consistency verification results, the multi-scale prediction output set is extracted;

[0097] Based on the multi-scale prediction output set, multiple predicted load stripes are constructed, wherein each predicted load strip includes an upper limit curve and a lower limit curve, and the time length of each predicted load strip corresponds one-to-one with the time length of multiple analysis windows.

[0098] The intersection load strip of multiple predicted load strips is taken and output as the predicted load information.

[0099] In this embodiment of the application, the multi-scale prediction output set that has passed the post-consistency verification is first extracted based on the post-consistency verification result.

[0100] Secondly, based on the multi-scale forecast output set, the fluctuation range of the predicted values ​​is calculated, and multiple predicted load stripes are constructed. Each predicted load strip includes an upper limit curve and a lower limit curve, and the time lengths of the multiple predicted load strips correspond one-to-one with the time lengths of multiple analysis windows. For example, a 1-hour analysis window corresponds to a 1-hour strip, and a 1-day analysis window corresponds to a 1-day strip. The predicted load stripes reflect the reasonable fluctuation range of the predicted values ​​at that time scale. For example, for the forecast results in each multi-scale forecast output set, the fluctuation range of the predicted values ​​can be calculated through statistical analysis: first, all predicted values ​​at each time point in the multi-scale forecast output set are extracted; then, based on a preset confidence level (e.g., 95% confidence interval), the upper quantile (e.g., 95th quantile) of the predicted value at the corresponding time point is taken as the value of the upper limit curve, and the lower quantile (e.g., 5th quantile) is taken as the value of the lower limit curve, thus constructing the predicted load strip at that time scale.

[0101] Finally, the intersection load strip of multiple predicted load strips is taken and output as the predicted load information. The intersection load strip is the area simultaneously covered by all predicted load strips, ensuring that the prediction results are within a reasonable range at all time scales. For example, for all predicted load stripes, the intersection load strip is calculated point by point along the time axis: for each time point, the minimum upper limit of all stripes at that time point is taken as the upper limit of the intersection strip, and the maximum lower limit of all stripes at that time point is taken as the lower limit of the intersection strip, forming an intersection area that only includes the common coverage of all stripes. For example, if there are predicted load stripes with three time lengths of 1 hour, 1 day, and 1 week, and the upper and lower limits of the 1-hour strip at a certain time point t1 are [12kW, 20kW], the 1-day strip is [15kW, 18kW], and the 1-week strip is [14kW, 19kW], then the intersection load strip at that time point is [15kW, 18kW]. By processing point by point in this way, a complete intersection load strip is finally formed, which can be used as the final predicted load information that takes into account the reliability of each scale.

[0102] In summary, compared to existing technologies, this application updates and collects real-time multi-source feature data of the target charging pile, and combines this real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction and obtain predicted load information. Thus, by updating and collecting real-time multi-source feature data, the latest load influencing factors can be captured. Combined with the multi-scale specialization capabilities of the multi-channel prediction component for joint prediction, the model achieves real-time dynamic adaptation and integrates the prediction advantages of different time scales, ensuring that the final output predicted load information is both reliable and accurate, and closely reflects actual scene changes.

[0103] In summary, the embodiments of this application have at least the following technical effects:

[0104] Compared to existing technologies, this application first collects multi-source feature data related to the load of the target charging pile, and then performs feature fusion on the multi-source feature data to form sample time-series data. In this way, it breaks through the information limitations of a single data source, transforms scattered influencing factors into structured and time-series input samples, and provides multi-source and reliable structured data for subsequent multi-scale analysis and model prediction.

[0105] Secondly, this application employs analysis windows of various time lengths to divide the sample time-series data into multiple scales, obtaining multi-scale sample time-series data. Based on the multi-scale sample time-series data, multiple long short-term memory network prediction channels corresponding to the analysis windows are constructed and trained to obtain multi-channel prediction components. This avoids the attentional distraction problem of a single model in multi-scale feature learning, providing multi-dimensional model support for subsequent joint load prediction using real-time data.

[0106] Finally, this application updates the real-time multi-source feature data of the target charging piles and combines the real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction and obtain predicted load information. In this way, by updating the real-time multi-source feature data, the latest load influencing factors can be captured. Combined with the multi-scale specialization capabilities of the multi-channel prediction component for joint prediction, the model adapts to real-time dynamics and integrates the prediction advantages of different time scales, ensuring that the final output predicted load information is both reliable and accurate, and closely reflects changes in the actual scenario.

[0107] Through the above technical solution, this application overcomes the shortcomings of traditional methods that rely on single data and lack feature dimensions by collecting multi-source feature data related to the target charging pile load. It can comprehensively capture multiple load driving factors, providing complete data support for prediction. By employing analysis windows of various time lengths to divide multi-scale sample time-series data and constructing and training corresponding long short-term memory network prediction channels to form a multi-channel prediction component, it overcomes the problem that single-time-scale models cannot adequately account for changes in trends over different periods. Furthermore, by updating and collecting real-time multi-source feature data and combining it with the multi-channel prediction component to conduct joint load prediction, it solves the problems of insufficient real-time performance and large deviations in handling dynamic scenarios inherent in traditional methods. In this way, it ensures that the charging pile load prediction results can adapt to dynamic factors such as changes in user behavior and sudden environmental impacts in real time, improving the accuracy and real-time performance of the prediction results. This provides accurate and reliable load data for grid-source-load-storage coordinated scheduling, charging pile resource allocation optimization, and improvement of user charging experience.

[0108] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0109] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0110] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0111] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0112] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0113] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0114] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A charging pile load prediction method based on LSTM and multi-source feature fusion, characterized in that, The method comprises the following steps: Collecting load-related multi-source feature data of the target charging pile and performing feature fusion on the multi-source feature data to form sample time series data; Using multiple time length analysis windows to perform multi-scale division on the sample time series data to obtain multi-scale sample time series data, and constructing and training multiple long short-term memory network prediction channels corresponding to the analysis windows according to the multi-scale sample time series data to obtain a multi-channel prediction component; Updating the real-time multi-source feature data of the target charging pile, and combining the real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction to obtain predicted load information; Wherein, updating the real-time multi-source feature data of the target charging pile, and combining the real-time multi-source feature data with the multi-channel prediction component to perform joint load prediction to obtain predicted load information, comprises: Using the longest analysis window as a collection constraint, updating the real-time multi-source record information of the target charging pile, and extracting the real-time multi-source feature data based on feature engineering; Based on the analysis window of multiple time lengths, the real-time multi-source feature data is divided into multiple scales to obtain a multi-scale prediction input set; Based on the multi-scale prediction input set, the multi-channel prediction component is verified for post-consistency, and if passed, the predicted load information is obtained by joint load fusion based on the post-consistency verification result; Wherein, based on the multi-scale prediction input set, the multi-channel prediction component is verified for post-consistency, comprising: According to the corresponding relationship between the analysis window and the multiple long short-term memory network prediction channels, the multi-scale prediction input set is input into the multiple long short-term memory network prediction channels and iteratively predicted for a continuous preset number of times to obtain a multi-scale prediction output set; Curve fitting is performed on each prediction output subset in the multi-scale prediction output set to obtain a corresponding typical output curve set; Based on the typical output curve set, random combination is performed, and the mutual residual error between the multiple typical output curves is calculated based on the combination result; If each mutual residual error meets a preset consistency threshold, it is determined that the post-consistency verification is passed, and joint load fusion is allowed.

2. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 1, wherein, Collecting load-related multi-source feature data of the target charging pile and performing feature fusion on the multi-source feature data to form sample time series data, comprising: Collecting multi-source record information of the target charging pile, including historical load sequence information, time period information, user behavior information and external environment information; Based on feature engineering and the multi-source record information, the multi-source feature data is extracted and spliced to obtain the sample time series data.

3. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 1, characterized in that, Using multiple time length analysis windows to perform multi-scale division on the sample time series data to obtain multi-scale sample time series data, comprising: Based on the historical load prediction record of the target charging pile, the time length of multiple historical analysis windows is extracted, and multiple time length analysis windows are defined correspondingly; According to the analysis window of a plurality of time lengths defined, the sample time series data is respectively processed by sliding segmentation to generate corresponding multi-scale sample time series data, wherein the multi-scale sample time series data comprises a plurality of sample time series data subsets.

4. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 3, characterized in that, According to the multi-scale sample time series data, a plurality of long short-term memory network prediction channels corresponding to the analysis window are respectively constructed and trained to obtain a multi-channel prediction component, comprising: Based on the analysis window of a plurality of time lengths, the same number of long short-term memory network prediction channels is constructed correspondingly; Taking a plurality of sample time series data subsets as training data, a plurality of long short-term memory network prediction channels are respectively trained, and the prediction residual of each long short-term memory network prediction channel is calculated correspondingly; When each prediction residual is less than a preset individual accuracy threshold, it is determined that the plurality of long short-term memory network prediction channels pass the pre-training verification, and are integrated and output as the multi-channel prediction component.

5. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 4, characterized in that, Based on the post-consistency verification result, the joint load fusion is performed to obtain the predicted load information, comprising: Based on the post-consistency verification result, the typical output curve set is extracted; According to the typical output curve set, segmented weighted fusion is performed, wherein the weighted weight of the typical output curve is monotonically positively correlated with the time length of the corresponding analysis window; The segmented weighted fusion result is output as the predicted load information.

6. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 5, characterized in that, According to the typical output curve set, segmented weighted fusion is performed, wherein the weighted weight of the typical output curve is monotonically positively correlated with the time length of the corresponding analysis window, comprising: Taking the window end point of each analysis window as a segmentation point, the typical output curve set is segmented to obtain a plurality of segmented curve subsets, wherein the number of segmented output curves included in each segmented subset decreases along the time direction; Iterating through a plurality of segmented curve subsets, the preset window weight corresponding to a plurality of analysis windows in each segmented curve subset is normalized respectively to obtain a plurality of segmented weighted weight sets; Based on a plurality of segmented weighted weight sets, a plurality of segmented curve subsets are weighted fused respectively to obtain the segmented weighted fusion result.

7. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 5, characterized in that, The segmented weighted fusion result is output as the predicted load information, further comprising: Based on the time sequence relationship, the segmented weighted fusion result is spliced, and the splicing breakpoints are identified; Combining a preset smoothing function, a plurality of splicing breakpoints are smoothed to define the smoothing result as the predicted load information.

8. The charging pile load prediction method based on LSTM and multi-source feature fusion according to claim 4, characterized in that, Based on the post-consistency verification result, the joint load fusion is performed to obtain the predicted load information, further comprising: Based on the post-consistency verification result, the multi-scale prediction output set is extracted; According to the multi-scale prediction output set, a plurality of prediction load strips are constructed, wherein the prediction load strip comprises an upper limit curve and a lower limit curve, and the time length of the plurality of prediction load strips corresponds one by one to the time length of the plurality of analysis windows; Taking the intersection load strip of a plurality of prediction load strips as an output, the predicted load information is output.

Images