Traction load short-term prediction method based on multi-dimensional features and sta-unet

Abstract

The application discloses a traction load short-term prediction method based on multi-dimensional features and STA-UNet, which comprises the following steps: obtaining an original data set of a traction substation; extracting traction load operation state features based on a sliding window; constructing traction load multi-dimensional features based on time, train operation diagrams and power data, and obtaining a feature set after fusion; inputting the feature set into a STA-UNet network with a soft threshold attention mechanism for training, and using the trained STA-UNet network to predict the load of a target traction substation. The input features constructed by the application fuse periodic time encoding and multi-dimensional information extracted from train operation diagrams, and the STA-UNet network model proposed by the application realizes efficient fusion of multi-scale time sequence features between an encoder and a decoder through a soft threshold attention mechanism and a skip connection structure, thereby improving the prediction accuracy of the traction load.

Description

Technical Field

[0001] This invention relates to the field of traction load prediction technology, and in particular to a method and system for short-term traction load prediction based on multidimensional train operation characteristics and STA-UNet. Background Technology

[0002] With the continuous expansion of railway operation scale and the electrification rate of high-speed railways, energy consumption has also increased significantly. Among them, traction energy consumption accounts for about 60%-70% of the total energy consumption of railways. Therefore, energy conservation and consumption reduction of electrified railways has become an important research direction.

[0003] Electric locomotives preferentially employ regenerative braking during braking, generating regenerative braking energy that accounts for approximately 10%-40% of the total traction energy consumption. However, existing research indicates that only about 50% of this energy is effectively utilized within the traction system, with the majority being fed back to the grid as low-quality electricity, failing to achieve effective storage and dispatch. Therefore, improving the utilization rate of regenerative braking energy is one of the key technological pathways to achieving green railways.

[0004] Introducing power flow control and energy storage devices into traction power supply systems to store, transfer, and reuse regenerative energy is a key technical means to improve the utilization rate of regenerative braking energy and reduce energy consumption. The effective control of these power electronic devices relies heavily on prior information provided by short-term traction load forecasts. Adjusting the system's operating state in advance based on high-precision forecasts ensures the real-time performance and effectiveness of regenerative energy management strategies. Therefore, achieving high-precision short-term traction load forecasts is a crucial prerequisite for improving the overall system energy efficiency and ensuring the safe and stable operation of the power grid. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for short-term traction load prediction based on multi-dimensional train operation characteristics and STA-UNet. By utilizing prior physical information to enhance the model's ability to capture the changing patterns of time-series data, this invention predicts the load of high-speed railway traction substations, thereby improving the accuracy of short-term traction load prediction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, this invention provides a short-term traction load prediction method based on multi-dimensional train operation characteristics and STA-UNet, comprising the following steps: Obtain the raw dataset of the traction substation; the raw dataset includes the train operation diagram and the active power data of the traction substation; Extract traction load operating status features based on a sliding window; Construct a multidimensional feature set of traction load based on time, train operation diagram and power data, including periodic encoding of traction load time series features, construction of multidimensional operation features based on train operation diagram and fusion of multidimensional features, and obtain a feature set after fusion; The feature set is input into the STA-UNet network based on the soft thresholding attention mechanism for training. The STA-UNet consists of three symmetrical encoder-decoder layers. The trained STA-UNet network is used to predict the load of the target traction substation and output the prediction result.

[0007] On the other hand, the present invention provides a short-term traction load prediction system based on multi-dimensional train operation characteristics and STA-UNet, comprising the following modules: Data acquisition module: used to acquire raw datasets and train operation diagrams from traction substations; Traction load operating status feature extraction module: used to extract traction load operating status features based on a sliding window; Feature set construction module: used to construct multi-dimensional features of traction load based on time, train operation diagram and power data, and obtain the feature set after fusion; Prediction module: The feature set is input into the STA-UNet network based on the soft thresholding attention mechanism for training, and the trained STA-UNet network is used to predict the load of the target traction substation and output the prediction results.

[0008] Compared with the prior art, the present invention has the following beneficial effects: To address the strong abrupt changes and non-stationarities of traction loads caused by the spatiotemporal superposition of multiple trains, this invention systematically extracts key operational elements such as train type, train number, train formation, and scheduled arrival and departure times from the train timetable. These elements are then deeply integrated with power sequences and periodic time codes to form a multi-dimensional input feature set that reflects the traction load mechanism. The input features constructed in this invention integrate multi-dimensional information extracted from periodic time codes and train timetables, enabling deep learning models to explicitly utilize the physical prior knowledge inherent in the operation plan, effectively enhancing the model's ability to express load abrupt changes and overall trends.

[0009] The STA-UNet network model proposed in this invention effectively suppresses noise features and highlights key mutation points through a soft-thresholding attention mechanism. At the same time, it achieves efficient fusion of multi-scale temporal features between the encoder and decoder based on a skip connection structure, thereby improving the prediction accuracy of traction load. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall process of Embodiment 1 of the present invention; Figure 2This is a schematic diagram of the STA-UNet network structure in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram showing the comparison between the predicted value and the actual value under the single-step prediction condition in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram showing the comparison between predicted and actual values ​​under the multi-step prediction conditions of Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the output results of Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the overall structure of Embodiment 2 of the present invention; Figure 7 This is a schematic diagram of the electronic device structure according to Embodiment 2 of the present invention. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Example 1: Please see Figures 1-5 A short-term traction load prediction method based on multi-dimensional train operation characteristics and STA-UNet, in a specific embodiment, includes the following steps: Step 1: Obtain the raw dataset of the traction substation, which includes the active power data of the traction substation. And train operation diagram.

[0013] Step 2: Extract traction load operating status features based on sliding window.

[0014] The following steps are adopted: 1) Data pruning: In the daily traction load curve, there is a long window period, and this part of the data has limited contribution to model learning and may even introduce noise. Therefore, the data with long window periods is pruned before building the model input.

[0015] 2) Window definition: Starting from the power within the window, select the power... ,in For window distance, For index The power sample value at that time.

[0016] 3) Data Summation: Take the absolute values ​​of the data within the window and sum them to obtain the total sum. .

[0017] 4) Window discrimination: When the cumulative result... Greater than ( This is the threshold scaling factor. If the power accumulation threshold is used as the reference power, then it is determined to be the operating condition, and... Label the corresponding data as 1; when the cumulative result... Less than or equal to If so, it is determined to be an unloaded operating condition, and The corresponding data is labeled 0; then, the sliding window moves forward one time step, and the above process is repeated until the entire data sequence is traversed, completing the labeling of load mutations within the day.

[0018] Step 3: Construct a multi-dimensional feature set of traction load based on time, train operation diagram and power data, and then fuse them to obtain the feature set.

[0019] Periodic encoding of traction load timing characteristics. To preserve local timing context information, this embodiment uses sine and cosine functions to periodically position the time. The unified periodic encoding is as follows: ; ; In the formula, A periodic position encoding function for time attributes, used to encode time steps. Time attribute Mapped to the encoded vector of the model input; The time attribute of the time step. , These correspond to hours, minutes, seconds, and days, respectively. Periods with different time granularities; For encoding dimension indexes.

[0020] By periodically encoding time, time information is mapped into features that the STA-UNet network model can learn. This allows the model to more fully capture the patterns of load changes over time, thereby improving prediction accuracy.

[0021] Construct multi-dimensional operational features based on train timetables. Extract multi-dimensional operational features such as train number, train formation, and train arrival and departure times from the train timetables, and combine them with the aforementioned time coding to construct sequence features based on the arrival and departure times of trains.

[0022] For each train Extracting discrete elements and event times from train operation diagrams:

[0023] In the formula For vehicle models, For train number, For grouping, , These are the train's arrival and departure times, respectively.

[0024] Then, using an embedding method, discrete features such as vehicle type, train number, and train formation are mapped into low-dimensional dense vectors: Establish embedding maps separately to map discrete vectors to dense vectors:

[0025] The discrete movements of the train are embedded and concatenated into a movement vector: =[ ] In the formula, An embedded encoding function for discrete vehicle type identifiers, used to map vehicle IDs to low-dimensional dense vectors; For the first k The model embedding vector of each train; For the first k Train number embedding vector; For the first k Embedded vectors of train formations; An embedded encoding function for discrete train number identifiers, used to map train number IDs to low-dimensional dense vectors; An embedded encoding function for grouping discrete identifiers, used to map train IDs to low-dimensional dense vectors.

[0026] Multidimensional feature fusion. First, the time and power data are aligned with a sampling step size of 1 second to obtain the power value at each time step. This forms a unified time series sample. Subsequently, a multi-channel input structure is used to introduce the aforementioned multi-dimensional feature information: Construct the final input vector: ] In the formula This is the concatenated vector.

[0027] The arrival and departure times of each train are extracted from the train operation schedule. When the time series reaches the arrival time of a certain train, the corresponding train operation features are loaded to achieve a time-series correspondence between train operation status and power data. Finally, all time, power, and operation features are concatenated and fused to form a complete input feature set.

[0028] Step 4: Input the feature set into the STA-UNet network with soft thresholding attention mechanism for training, and use the trained STA-UNet network to predict the load of the target traction substation.

[0029] Includes the following steps: The input feature set is divided into training, validation, and test sets in an 8:1:1 ratio. The processed power data is then numerically normalized using the max-min normalization method. ; In the formula, To train the minimum value of concentrated power, To train the maximum value of the concentrated power, for, For a moment The original power values. The pre-divided dataset is divided into multiple small sample subsets in batches to support multiple weight updates during the training process of the deep learning model.

[0030] The training set data is input into a UNet network (Soft-thresholdAttention U-Net, STA-UNet) with a soft-threshold attention mechanism for training. The model structure of STA-UNet is as follows: Figure 2 As shown, STA-UNet consists of three symmetrical encoder-decoder layers.

[0031] like Figure 2 As shown in the left half, the encoder includes a BiLSTM module, a perceptual network, an average pooling layer, and a soft-thresholding attention mechanism.

[0032] The BiLSTM module, while expanding the input feature dimension, can capture the forward and backward dependencies of the sequence, thus more comprehensively characterizing the temporal features of traction load. The calculation process is as follows: ; ; ; In the formula, For LSTM at time The forward hidden state vector; For LSTM at time The forward unit state vector; State update operator for Long Short-Term Memory (LSTM) networks; The input vector; For LSTM at time The backward hidden state vector; For LSTM at time The backward unit state vector; For bidirectional long-term and short-term networks at time The output vector is used to simultaneously represent the historical dependencies and future context information of the input sequence at that moment.

[0033] The perceptual network in the encoder is used to adjust the data dimensionality to capture the spatiotemporal correlation of traction load data; the average pooling layer is used to reduce the dimensionality of time-series data while expanding the receptive domain of the perceptual network; the soft-thresholding attention mechanism is used to suppress irrelevant or noisy feature components and highlight information highly relevant to the prediction task. The soft-thresholding attention mechanism first performs a sign function operation on the low-level features to determine their importance; then, it weights and adjusts the features based on a threshold coefficient, thereby enhancing the network model's responsiveness to key features. The calculation expression is: ; In the formula, For a moment The gated output vector; For symbolic functions, The value is 1 if the condition is met, otherwise it is 0. For a moment The input signal; This is the threshold parameter, which is learned autonomously by the model. This is the maximum value operator, used to select the larger of two candidate values.

[0034] like Figure 2 As shown in the right half, each decoder layer has the data size and dimension of its corresponding encoder layer. The decoder includes a perceptual network and a skip connection network; the perceptual network in the decoder is used to perform dimensionality transformation on the input features, thereby recovering and reconstructing high-level feature information; the skip connection network is used to fuse low-resolution features extracted from the encoder with high-resolution features of the same size generated in the decoder. Since features from different levels are usually simply concatenated, assuming equal importance, this does not match the actual feature distribution. To adaptively match and filter the feature channels of the encoder and decoder, an improved skip connection mechanism is established, the calculation expression of which is: ; ; These are adaptive weighting coefficients; This is the sigmoid function, used to map output values ​​to the [0,1] interval; This is the weight parameter matrix; Output features for the encoding branch; Output features for the decoding branch; For bias term parameters; This is the intermediate feature vector after fusion.

[0035] The improved skip connection mechanism can effectively retain the important features selected by the encoder and pass them to the corresponding layer of the decoder, thereby significantly enhancing the model's ability to learn and extract the spatiotemporal features of traction load.

[0036] In the STA-UNet network, the encoder performs layer-by-layer feature extraction and downsampling on the input features to obtain temporal representations at different scales. The decoder uses an upsampling module to restore the feature resolution step by step and fuses it with the corresponding scale features from the encoder through skip connections, thereby preserving both global trend information and local abrupt changes. Specifically, the upsampling module uses interpolation upsampling to map low-resolution features to a high-resolution feature space. To enhance nonlinear expressive power, the module introduces the Tanh activation function to perform nonlinear transformation on the features. A fully connected layer is set at the end of the network to map the fused high-dimensional feature vector to the target-dimensional load prediction output, thereby enabling single-step or multi-step prediction of the target traction substation load.

[0037] In each training iteration, the STA-UNet network model first performs forward propagation on the input samples, outputting the corresponding load prediction value. Then, it compares the predicted value with the actual load value at the same time step, using the mean squared error (MSE) as the loss function to measure the prediction error. After obtaining the loss value, it uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the trainable parameters of each layer of the network, and then uses the Adam optimizer to update the parameters to achieve model optimization. The aforementioned process of "forward propagation—loss calculation—backward propagation—parameter update" is executed batch-by-batch on the training set, and the model is evaluated using a validation set after each training round to monitor the model's convergence status. This training process is repeated until the preset number of training rounds is reached or the loss function meets the convergence condition, thus obtaining the trained STA-UNet load prediction model.

[0038] Then, the STA-UNet network model is validated and tested using validation and test sets. Finally, the trained STA-UNet network is used to predict the load of the target traction substation, and the predicted values ​​are output. The output results are as follows: Figure 5 As shown.

[0039] like Figures 3 to 4 As shown, the STA-UNet model proposed in this invention can more accurately follow the load change trend and has a better fitting ability for load abrupt changes and fluctuations, thereby achieving higher prediction accuracy.

[0040] Example 2 like Figure 6 As shown, a short-term traction load prediction system based on multidimensional features and STA-UNet, using the above method, includes the following modules: Data acquisition module: used to acquire raw datasets and train operation diagrams from traction substations; Traction load operating status feature extraction module: used to extract traction load operating status features based on a sliding window; Feature set construction module: used to construct multi-dimensional features of traction load based on time, train operation diagram and power data, and obtain the feature set after fusion; Prediction module: The feature set is input into the STA-UNet network with soft thresholding attention mechanism for training, and the trained STA-UNet network is used to predict the load of the target traction substation.

[0041] The traction load short-term prediction system based on multidimensional features and STA-UNet can be installed in computer equipment.

[0042] The module described in this invention refers to a series of computer program segments that can be executed by the processor of a computer device and can perform a fixed function, and which are stored in the memory of the computer device.

[0043] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment, and the corresponding content in the aforementioned method embodiment can be referred to.

[0044] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0045] Figure 7 A block diagram is shown that is suitable for implementing embodiments of the present application. Figure 7 The electronic device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0046] like Figure 7 As shown, the electronic device is represented in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors 410, memory 430, and communication bus 440 connecting different system components (including memory 430 and processing unit 410).

[0047] Communication bus 440 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MAC) buses, Enhanced ISA buses, Video Electronics Standards Association (VESA) local buses, and Peripheral Component Interconnect (PCI) buses.

[0048] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0049] Memory 430 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The electronic device may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 430 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application.

[0050] A program / utility having a set (at least one) of program modules can be stored in memory 430. Such program modules include—but are not limited to—an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The program modules typically perform the functions and / or methods described in the embodiments of this application.

[0051] Processor 410 executes various functional applications and data processing by running programs stored in memory 430, such as implementing embodiments of this application. Figure 1 The embodiment shown provides a short-term traction load prediction method based on multidimensional features and STA-UNet.

[0052] This application provides a non-transitory computer-readable storage medium that stores computer instructions, which cause the computer to execute embodiments of this application. Figure 1 The embodiment shown provides a short-term traction load prediction method based on multidimensional features and STA-UNet.

[0053] The aforementioned computer-readable storage medium may be any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium that contains or stores a program that may be used by or in connection with an instruction execution system, apparatus, or device.

[0054] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0055] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0056] Computer program code for performing the operations of the embodiments of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed 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 cases involving remote computers, the remote computer can 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 it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0057] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0058] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A short-term traction load prediction method based on multidimensional features and STA-UNet, characterized in that, Includes the following steps: Obtain the raw dataset from the traction substation; Extract traction load operating status features based on a sliding window; A multi-dimensional feature set of traction load based on time, train operation diagram and power data is constructed and fused to obtain the feature set. The feature set is input into the STA-UNet network based on the soft thresholding attention mechanism for training, and the trained STA-UNet network is used to predict the load of the target traction substation and output the prediction results.

2. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 1, characterized in that, Extracting traction load operating state features based on a sliding window includes the following steps: Data cropping: cropping out data from long periods of inactivity; Define window: Starting from the power within the window, select the power... ; Data Summation: The absolute values ​​of the data within the window are summed to obtain the total sum. ; Window discrimination: when the sum of results Greater than If so, it is determined to be an operating condition, and Label the corresponding data as 1; when the cumulative result... Less than or equal to If so, it is determined to be an unloaded operating condition, and Label the corresponding data with 0; traverse the entire data sequence to complete the labeling of load mutations within the day; In the formula, For index Power sampling value at time; This is the threshold scaling factor; The base power accumulation threshold.

3. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 1, characterized in that, Constructing a multidimensional feature of traction load based on time, train operation diagrams, and power data includes the following steps: Sine and cosine functions are used to periodically encode the time position to capture the load variation over time, thereby obtaining the traction load time sequence characteristics; Constructing multi-dimensional operational features based on train operation diagrams; The dataset is obtained by splicing and fusing time, power, and multi-dimensional operational features.

4. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 3, characterized in that, The calculation expression for periodic position encoding of time is as follows: ； ； In the formula, A periodic position encoding function for time attributes, used to encode time steps. Time attribute Mapped to the encoded vector of the model input; The time attribute of the time step. , These correspond to hours, minutes, seconds, and days, respectively. Periods with different time granularities; For encoding dimension indexes.

5. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 4, characterized in that, The method for constructing multi-dimensional operational features based on train operation diagrams is as follows: The train arrival and departure times extracted from the train operation diagram are combined with the traction load time series features. Sequence features are constructed through the train arrival and departure times, and then the discrete features of the train are mapped into low-dimensional dense vectors.

6. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 3, characterized in that, The splicing and fusion process includes the following steps: First, the time and power data are aligned according to the sampling step size to form a unified time series sample; Then, based on the train's arrival and departure times, when the train's arrival time arrives, the corresponding train operation characteristics are loaded to achieve a time-series correspondence between train operation status and power data. Finally, all time, power, and train operation characteristics are spliced ​​and fused to obtain a feature set.

7. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 1, characterized in that, The STA-UNet network consists of a three-layer symmetrical encoder and decoder. The encoder includes a BiLSTM module, a perceptual network, an average pooling layer, and a soft thresholding attention mechanism. The decoder includes a perceptual network and a skip connection network.

8. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 7, characterized in that, The BiLSTM module is used to capture the forward and backward dependencies of the sequence; the perceptual network in the encoder is used to adjust the data dimension to capture the spatiotemporal correlation of traction load data; the average pooling layer is used to reduce the dimension of time-series data while expanding the receptive field of the perceptual network; the soft thresholding attention mechanism is used to suppress irrelevant or noisy feature components and highlight information that is highly relevant to the prediction task. The perceptual network in the decoder is used to perform dimensionality changes on the input features, thereby recovering and reconstructing high-level feature information; The skip connection network is used to fuse low-resolution features extracted from the encoder with high-resolution features of the same size generated in the decoder.

9. The method for short-term traction load prediction based on multidimensional features and STA-UNet according to claim 8, characterized in that, The computational expression for the soft-thresholding attention mechanism is: ； The working process of the hop connection network is as follows: ； ； In the formula, For a moment The gated output vector; For symbolic functions, The value is 1 if the condition is met, otherwise it is 0. For a moment The input signal; This is the threshold parameter, which is learned autonomously by the model. For finding the maximum value operator; These are adaptive weighting coefficients; This is the sigmoid function, used to map output values ​​to the [0,1] interval; This is the weight parameter matrix; Output features for the encoding branch; Output features for the decoding branch; For bias term parameters; This is the intermediate feature vector after fusion.

10. A short-term traction load prediction system based on multidimensional features and STA-UNet, using the method described in any one of claims 1-9, characterized in that, Includes the following modules: Data acquisition module: used to acquire raw datasets and train operation diagrams from traction substations; Traction load operating status feature extraction module: used to extract traction load operating status features based on a sliding window; Feature set construction module: used to construct multi-dimensional features of traction load based on time, train operation diagram and power data, and obtain the feature set after fusion; Prediction module: The feature set is input into the STA-UNet network based on the soft thresholding attention mechanism for training, and the trained STA-UNet network is used to predict the load of the target traction substation and output the prediction results.

Images