Scheduling cost-oriented multi-representation fusion time sequence prediction method

Abstract

The invention discloses a scheduling cost-oriented multi-representation fusion time sequence prediction method, and belongs to the technical field of time sequence data processing. The method aims at solving the technical problems that an existing prediction model is insufficient in extreme value prediction, and the optimization target is disjointed from the actual scheduling cost. The method is technically characterized by comprising the following steps of: converting numeric time series data into time series image representation, and performing fusion modeling on the numeric representation and the image representation to capture a local form and a global trend of the data; and a scheduling cost perception loss function is adopted for model training, and the loss function can apply asymmetric cost penalty to under-prediction and over-prediction. According to the method, the prediction precision of sudden change areas such as peak values is improved through multi-representation fusion, and the prediction result more meeting the actual scheduling requirement is generated through the loss function guide model aligned with the service cost, so that the comprehensive scheduling cost and the service default risk are effectively reduced.

Description

A Multi-Representation Fusion Temporal Prediction Method Oriented to Scheduling Cost Technical Field

[0001] This invention relates to the field of time series data processing technology, and in particular to a multi-representation fusion time series prediction method oriented towards scheduling costs. Background Technology

[0002] With the rapid development of information technology and artificial intelligence, time-series forecasting technology plays a core role in intelligent scheduling systems in many key areas, such as load scheduling in power systems, resource allocation in cloud computing and edge computing, and traffic flow planning in intelligent transportation networks. These systems generally adopt a "predict first, then schedule" operating paradigm, that is, first predict key future indicators (such as grid load, server bandwidth, and traffic flow), and then formulate resource allocation and scheduling strategies based on the prediction results.

[0003] However, existing time-series forecasting methods for scheduling generally suffer from a core technical problem: there is a significant disconnect between the optimization objective of the forecasting model and the actual scheduling costs of business operations. On the one hand, existing forecasting models mostly rely on single numerical time-series representations for modeling. While they can capture the overall trend of the data, their prediction accuracy is often insufficient when facing non-stationary abrupt changes such as peaks and troughs in business traffic, easily leading to large prediction errors. On the other hand, these models are typically trained using symmetric loss functions such as mean squared error (MSE). These loss functions impose the same penalty on both over-prediction (predicted values ​​higher than the actual values) and under-prediction (predicted values ​​lower than the actual values) of equal magnitude. However, in actual scheduling scenarios, the business costs resulting from these two types of prediction errors are highly asymmetrical. For example, in edge computing, under-prediction may lead to insufficient bandwidth, service quality of service (SLA) defaults, or even service interruptions, with losses far exceeding the small resource redundancy costs caused by over-prediction. This mismatch between the prediction target and the scheduling cost means that even if the model performs well in terms of mathematical indicators, its prediction results are difficult to effectively guide scheduling decisions, often leading to problems such as unreasonable resource allocation, high risk of SLA default, and waste of operating costs. Summary of the Invention

[0004] The main objective of this invention is to provide a multi-representation fusion time-series prediction method oriented towards scheduling costs, in order to solve the problems raised in related technologies.

[0005] To achieve the above objectives, according to one aspect of the present invention, a multi-representation fusion temporal prediction method oriented towards scheduling cost is provided, comprising: converting numerical time-series data into a time-series image representation; extracting numerical features from the numerical time-series data and image features from the time-series image representation respectively; fusing the numerical features and the image features to obtain a hybrid representation; generating a predicted value based on the hybrid representation; and training the model using a scheduling cost-aware loss function, wherein the scheduling cost-aware loss function is used to impose an asymmetric prediction cost on under-prediction and over-prediction.

[0006] As a preferred embodiment of the present invention, the step of converting numerical time-series data into a time-series image representation includes: normalizing the numerical time-series data; mapping the data points on the normalized time-series data to pixel coordinates on a canvas; connecting the pixel coordinates in chronological order and rendering them as a polyline to generate the time-series image representation.

[0007] As a preferred embodiment of the present invention, the step of extracting image features from the temporal image representation includes: using a multi-scale feature extraction unit to perform convolution operations on the temporal image representation at at least two different scales to obtain multi-scale morphological feature maps; and integrating the multi-scale morphological feature maps to generate the image features.

[0008] As a preferred embodiment of the present invention, the step of extracting the numerical features of the numerical time series data includes: encoding the numerical time series data to obtain data dependency features; extracting structured time fields from the timestamps corresponding to the numerical time series data and encoding them to obtain periodic time features; and fusing the data dependency features and the periodic time features to obtain the numerical features.

[0009] As a preferred technical solution of the present invention, after fusing the numerical features and the image features, the method further includes: processing the hybrid representation using a cross-representation dual-axis dependency modeling structure to capture the dependencies in the representation dimension and the sequence dimension.

[0010] As a preferred embodiment of the present invention, the cross-representation biaxial dependency modeling structure includes: a representation mixer for transforming along the representation dimension of the mixed representation to model the interaction between the representations; and a sequence mixer for transforming along the sequence dimension of the mixed representation to model the temporal dependency.

[0011] As a preferred embodiment of the present invention, the scheduling cost-aware loss function includes: an asymmetric quantile loss term, used to apply an asymmetric linear penalty to the prediction error according to a preset ratio of underprediction cost to overprediction cost; and a soft threshold term, used to apply a smooth, simulated one-time penalty when the underprediction error exceeds a preset tolerance band.

[0012] As a preferred embodiment of the present invention, the soft threshold is implemented by a parameterized sigmoid function, wherein the parameter is used to control the intensity of the penalty, the tolerance band for triggering the penalty, and the smoothness of the threshold.

[0013] A time series prediction system includes: a representation conversion module for converting input numerical time series data into time series image representations; a feature extraction module for extracting numerical features from the numerical time series data and image features from the time series image representations; a representation fusion module for fusing the numerical features and the image features to obtain a hybrid representation; and a prediction and training module for generating predicted values ​​based on the hybrid representations and training the model using a scheduling cost-aware loss function, wherein the scheduling cost-aware loss function applies an asymmetric prediction cost to under-prediction and over-prediction.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0015] This invention provides a multi-representation fusion temporal prediction method oriented towards scheduling costs. By fusing numerical time-series data with its visual representations, this method fully utilizes the complementary information of numerical dependence and visual morphology. This method significantly enhances the model's ability to capture and predict morphological features in extreme regions such as peaks and valleys and non-stationary periods in time-series data, thereby improving prediction accuracy at key time points.

[0016] Furthermore, this invention constructs a scheduling cost-aware loss function, imposing asymmetric cost penalties on under-prediction and over-prediction during the model training phase, and smoothly integrating step costs, similar to one-time fines, from business rules into the optimization objective. This design ensures that the model training process no longer merely pursues the minimization of mathematical errors, but directly aligns with actual business scheduling costs, thereby guiding the model to generate prediction results that better meet economic efficiency and risk aversion requirements, effectively reducing the risk of Service Level Agreement (SLA) default and overall scheduling costs caused by under-prediction. Attached Figure Description

[0017] Figure 1 is a schematic diagram of the overall process of a multi-representation fusion temporal prediction method oriented towards scheduling cost provided by an embodiment of the present invention.

[0018] Figure 2 is a schematic diagram of the structure of the time series numerical feature extraction part in one embodiment of the present invention.

[0019] Figure 3 is a schematic diagram of the temporal image feature extraction module in one embodiment of the present invention.

[0020] Figure 4 is a schematic diagram of the cross-representation biaxial dependency modeling module in one embodiment of the present invention.

[0021] Figure 5 is a comparative schematic diagram of the prediction effect in one embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this embodiment can be combined with each other. 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 a part of the embodiments of the present invention, and not all of them. 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.

[0023] This invention provides a multi-representation fusion temporal prediction method for scheduling costs. This method aims to solve the technical problems of insufficient prediction of extreme / abrupt scenarios in the existing technology for time series prediction, and the disconnect between the prediction target and the actual scheduling cost.

[0024] Referring to Figure 1, Figure 1 illustrates an end-to-end flowchart of a scheduling cost-oriented multi-representation fusion temporal prediction method according to an embodiment of the present invention. This method can be applied to computer-implemented systems, such as those deployed on servers or cloud computing platforms. The method mainly includes the following steps: data acquisition and cleaning, representation transformation, hybrid representation learning, and model training and prediction. Specifically, the hybrid representation learning step jointly models both numerical and image representations within a unified framework, and supervises and optimizes this model through a scheduling cost-aware loss function, ultimately outputting high-precision prediction results aligned with business objectives.

[0025] The following will explain each step in detail.

[0026] Step 1: Data Acquisition and Cleaning. This step is responsible for acquiring raw time-series data and preprocessing it to provide a high-quality data foundation for subsequent modeling.

[0027] In a specific application scenario, such as traffic scheduling in edge computing, the collected data can include, but is not limited to, time-series data such as business traffic, service bandwidth load, and the number of user requests. The sampling granularity of this data can be at the minute, hour, or day level. Let the feature set after collection and preliminary processing be denoted as... ,in The length of the original sequence. The number of features.

[0028] To ensure the stability and accuracy of subsequent modeling, the collected raw data needs to be cleaned. In one embodiment, the data cleaning process is as follows: 1. Missing value handling: Verify the integrity of key fields (such as timestamps). If a timestamp is found to be missing, it can be filled with historical values ​​from the same period (e.g., the same time of the previous day) to maintain the continuity and periodicity of the sequence.

[0029] 2. Outlier Detection: Statistical methods (such as Z-score) are used to detect feature values. Specifically, the Z-score of each data point is calculated, and if its absolute value exceeds a preset threshold (e.g., 3), it is identified as an outlier. Outliers can be smoothed or replaced with neighboring values.

[0030] 3. Duplicate Record Handling: Check if there are duplicate data records with the same timestamp. If so, keep only one record and delete the rest.

[0031] After the above cleaning process, stable and clean numerical time series data are obtained for subsequent processing.

[0032] Step Two: Representation Transformation This step aims to transform the cleaned numerical time-series data... Transform into a visual representation This allows us to use visual features to capture information such as local morphology, trends, and abrupt changes in numerical sequences that are difficult to model explicitly. This step constitutes a configurable rendering pipeline.

[0033] In one embodiment, the representation transformation process includes the following sub-steps: Sub-step 2.1: Canvas initialization for Each one-dimensional time sequence in the features (in (For the input sequence length), initialize a fixed-size RGB three-channel canvas. Its size is For example, you can set the canvas width. With sequence length Equal, height Set the resolution based on experience (e.g., 200 pixels). The canvas background can be filled with a single color (e.g., white), and a coordinate system can be set.

[0034] Sub-step 2.2: Curve plotting: For each feature sequence Normalization is performed to map the numerical range to the canvas's coordinate space. Preferably, min-max normalization is used. ;in, For sequence In the The value of each time step.

[0035] Then, each point on the normalized sequence Pixel coordinates mapped to the canvas : ; Arrange adjacent coordinate points in chronological order. Connect them into a polyline and render it onto the canvas. Above. Rendering parameters are configurable to enhance the distinctiveness of visual features. For example: line width: can be set from 1 to 4 pixels and can adaptively adjust according to canvas resolution.

[0036] Color: You can set a uniform color for all features (e.g., black, RGB=(0,0,0)), or specify different colors for different feature dimensions.

[0037] Line style: You can choose from solid lines, dashed lines, etc.

[0038] Specific plotting operations can be implemented using mature image processing libraries. For example, the cv2.line() function from the OpenCV library or the plt.plot() function from the Matplotlib library can be used to draw polylines.

[0039] Example illustration: Assuming the input is a traffic flow sequence sampled every 5 minutes throughout the day, then the length... A single image can be initialized. The canvas is used to normalize the traffic sequence and then draw a 2-pixel wide black solid line on the canvas, ultimately generating a curve that visually reflects the daily traffic changes.

[0040] Sub-step 2.3: Resolution Adjustment: To improve the detail in key areas of the image (such as peaks and valleys), the generated canvas can be enlarged. Define an expansion factor. (Typically, the value ranges from 1.0 to 4.0), changing the canvas size from... Adjusted to : ; This operation can be accomplished using standard image scaling algorithms (such as bilinear interpolation), for example, by calling the OpenCV function `cv2.resize()`. The expanded image can more smoothly and clearly display the local fluctuation features of the sequence.

[0041] Example illustration: Based on the above Taking the canvas as an example, if the expansion factor is set... Then the canvas size is adjusted to The magnified image makes the transition details of the peaks and troughs in the flow more clear, which helps the visual feature extraction module capture this key information.

[0042] Sub-step 2.4: Channel initialization and storage: ... Each feature sequence is used to render the generated RGB image, which is then stacked along the channel dimension to form a multi-channel temporal image tensor. The final output image representation is as follows: This operation can be achieved using functions such as NumPy's np.stack() or PyTorch's torch.cat().

[0043] Example illustration: If the system simultaneously processes two characteristic sequences: bandwidth and traffic ( ), and the rendering canvas size is Then two images will be generated respectively. The RGB images are concatenated along the channel dimension to obtain the final image representation tensor. The bandwidth curves for channels 0-2 and the traffic curves for channels 3-5 are shown in the diagram.

[0044] Step 3: Hybrid Representation Learning. This step is the core of this invention and is responsible for the learning of numerical representations. Image representation generated in step two Joint modeling is performed to achieve information complementarity and enhancement. This step mainly includes temporal numerical feature extraction, temporal image feature extraction, representation fusion, and cross-representation biaxial dependency modeling.

[0045] Sub-step 3.1: Time-series numerical feature extraction: Referring to Figure 2, this module aims to extract periodicity, trend, and other dependent features from the original numerical sequence and its corresponding timestamps. The input includes the numerical sequence. and its corresponding timestamp sequence .

[0046] First, the numerical sequence The input is fed into a numerical feature encoding module, which preferably employs a one-dimensional convolutional network (Conv1d), but can also be a temporal convolutional network (TCN) or a multilayer perceptron (MLP). In one embodiment, a one-dimensional convolutional pair is used. Encoding is performed to capture local dependencies and obtain a high-dimensional numerical representation. : ;in, Indicates will The transpose is used to meet the input format requirements of Conv1d. The parameters of this Conv1d layer can be set as: number of input channels. Number of output channels The kernel size is 3 and the stride is 1.

[0047] At the same time, for timestamp sequences The process involves decomposing and extracting structured time fields such as year, month, week, day, hour, and minute. These fields explicitly provide periodic information to the model. In one embodiment, timestamp information is split and normalized into four-dimensional fields (e.g., hour / 24, day / 31, week / 7, month / 12) to form a time input tensor. Then it is mapped to the same dimension as the numerical representation through a linear layer. Time characteristics were obtained. : This linear layer has an input dimension of 4 and an output dimension of 256, and does not use a bias term. A Dropout layer (e.g., with a deactivation probability of 0.1) can be added afterward to prevent overfitting.

[0048] Finally, the numerical features With time characteristics The data is then fused. Fusion methods can include element-wise addition, concatenation, or attention mechanisms. In a preferred embodiment, element-wise addition is used, along with layer normalization and Dropout, to obtain the final time-series numerical features. : Sub-step 3.2: Temporal image feature extraction: Referring to Figure 3, this module is used to extract features from temporal images. Morphological and structural features are extracted. To balance multiple periodic patterns in long sequences with computational efficiency, this embodiment employs a multi-scale feature extraction unit.

[0049] The input is a time-series image tensor. This module first uses multi-scale convolutions...

[0050] Features are extracted at different scales. For each scale, the convolution operation is as follows: In one embodiment, three scales are used ( For example, different kernel sizes or strides can be used to achieve multi-scale operations. A more efficient approach is to use the same kernel with different strides. For example, a kernel size of [missing value]... The step size can be set to This yields feature maps with three different spatial resolutions. This image slicing-like operation can effectively reduce the computational complexity of subsequent processing while preserving key local information.

[0051] Subsequently, the multi-scale feature maps The data is fed into an image feature encoding module for information exchange and unification. In one embodiment, the IFE module performs the following operations: First, it upsamples (e.g., bilinear interpolation) the smaller-sized feature map (...). The resolution of ) is adjusted to match that of the largest feature map ( Then, all aligned feature maps are concatenated along the channel dimension. Finally, through a... Linear projection is performed on the convolutional layers to compress the stitched high-dimensional channels (e.g., 768 dimensions) to a uniform feature dimension. (e.g., 256 dimensions) to obtain image features .

[0052] To align with numerical features, The spatial dimensions are flattened and transposed to obtain the final temporal image features. ,in The length of the image feature sequence.

[0053] Sub-step 3.3: Representation Fusion: This sub-step integrates numerical features Image features Integration is then required. First, the two feature sequences need to be aligned to the same length along the time dimension. This can be achieved by downsampling longer sequences (e.g., average pooling) or upsampling shorter sequences (e.g., interpolation), denoted as... operate.

[0054] After alignment, a combination of concatenation and linear projection is used for fusion. The aligned numerical and image features are concatenated along the channel dimension, then projected through a linear layer, and finally regularization operations (such as GELU activation function and LayerNorm) are applied to obtain a unified hybrid representation. : ; ; Sub-step 3.4: Cross-representation biaxial dependency modeling: Referring to Figure 4, for feature fusion... To further capture complex temporal dependencies and cross-representation interactions, this invention designs a cross-representation dual-axis dependency modeling structure. This structure comprises two core components: a representation mixer (RepMix) and a sequence mixer (SeqMix).

[0055] The process flow of a single processing unit in this module is as follows: 1. First, process the input fused features. (Initial input is) Perform layer normalization.

[0056] 2. RepMix: Transforms along the representation (channel) dimension, i.e., for each time step... The dimensional feature vectors are processed independently. This is used to model the interactions between different representation sources (numerical, temporal, visual). In one embodiment, RepMix is ​​implemented by a multilayer perceptron (MLP) with the following structure: linear layer (D->4D)->GELU-> linear layer (4D->8D)->GELU->Dropout-> linear layer (8D->4D)->GELU-> linear layer (4D->D)->Dropout.

[0057] 1. Sequence Mixer (SeqMix): Performs transformations along the sequence dimension, i.e., for each channel... The time series are processed independently. This is used to model long-range or short-range dependencies between time steps. In one embodiment, SeqMix is ​​implemented by a multi-head self-attention mechanism with 4 heads, each head having a dimension of 64.

[0058] 2. The outputs of RepMix and SeqMix are superimposed and connected to the original input via residual concatenation. They are added together to ensure the stability of the training.

[0059] The above units can be stacked. Next (in one embodiment, The value is set to 2 (typically ranging from 1 to 10) to enhance the model's representational ability. The final output is then processed... Features after the second iteration .

[0060] Step 4: Establishing a Scheduling Cost-Aware Loss Function. To align the model's prediction target with the actual scheduling cost of the business, this invention proposes a scheduling cost-aware loss function. This loss function consists of two parts: a quantity-based asymmetric loss and a lightweight soft threshold term.

[0061] Set at time step The actual value is The predicted value is The error is Let the unit resource cost be... The unit revenue loss or opportunity cost caused by under-forecasting is .

[0062] First, define asymmetric quantile loss. Used to characterize the asymmetric cost of quantity-based computation: ;

[0063] When the cost of under-prediction Greater than the cost of over-prediction hour, The loss function imposes a greater penalty on under-prediction, leading the model to conservatively give slightly higher predictions.

[0064] Secondly, a soft threshold item was designed for the one-time fixed penalty triggered by unforeseen circumstances in simulated business operations (such as violation of SLA). This term applies when the underprediction error exceeds a certain tolerance band. It is activated at that time.

[0065] ;in: This refers to the intensity of the penalty per instance, the value of which can be determined based on the average fixed penalty in the business. In an embodiment oriented towards a cloud computing scenario, it is set to 4.

[0066] This is the tolerance range, i.e., the allowable under-forecasting range. It can be set according to business rules, for example, 2% to 5% of the demand.

[0067] It is a smoothing coefficient that controls the steepness of the threshold. The larger the value, the closer it is to the hard threshold. The typical value range is [range missing].

[0068] Ultimately, the total scheduling cost-aware loss function is the sum of the above two terms, averaged over all time steps: ;in It is the length of the predicted sequence.

[0069] Step 5: Model Training and Prediction The method of this invention is implemented based on a deep learning framework (such as PyTorch) and is trained and inferred on a computing device with GPU acceleration capabilities.

[0070] The training data is constructed using a sliding window approach. Given a long time series dataset, the length of the input window is set. and target prediction window length For example, it is advisable to... (Data for the 7 days prior to the forecast) (Data predicted for the next day).

[0071] The training process employs mini-batch stochastic gradient descent. In one embodiment, the Adam optimizer is used, and the batch size can be set to 32. During training, data from the validation set... The metrics are monitored, and an early stopping mechanism is implemented. For example, if the validation set loss does not improve within 10 consecutive evaluation periods, training is stopped to prevent overfitting.

[0072] Once trained, the model can be used to perform future calculations on new input sequences. The predictions are made at each time step, and the results can be directly used in downstream resource scheduling systems.

[0073] System Embodiments This embodiment of the invention also provides a time series prediction system that can execute the above-described method. The system may include: a data acquisition module for acquiring and cleaning time series data.

[0074] Representation Transformation Module: Used to render cleaned numerical time-series data into time-series image representations.

[0075] Hybrid Representation Learning Module: Used for joint learning of numerical representations and image representations. It can be further divided into numerical feature extraction unit, image feature extraction unit, representation fusion unit, and cross-representation biaxial dependency modeling unit.

[0076] Model training module: Used to optimize and train the parameters in the hybrid representation learning module using the scheduling cost-aware loss function.

[0077] Prediction module: Used to take new input data and generate future time series prediction results using a trained model.

[0078] The above system can be deployed on a single server, a distributed computing cluster, or a cloud platform. Its modules can be implemented through software, hardware, or a combination of both. These are merely preferred embodiments of the present invention and are not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make some modifications or alterations to the disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the scope of the present invention.

Claims

1. A multi-representation fusion temporal prediction method oriented towards scheduling costs, characterized in that, include: The numerical time-series data is converted into a time-series image representation; the numerical features of the numerical time-series data and the image features of the time-series image representation are extracted respectively. The numerical features and the image features are fused to obtain a hybrid representation; a predicted value is generated based on the hybrid representation, and the model is trained using a scheduling cost-aware loss function, which is used to impose an asymmetric prediction cost on under-prediction and over-prediction.

2. The method according to claim 1, characterized in that, The step of converting numerical time-series data into a time-series image representation includes: normalizing the numerical time-series data; mapping the data points on the normalized time-series data to pixel coordinates on the canvas; connecting the pixel coordinates in chronological order and rendering them as a polyline to generate the time-series image representation.

3. The method according to claim 1, characterized in that, The step of extracting image features from the temporal image representation includes: using a multi-scale feature extraction unit to perform convolution operations on the temporal image representation at at least two different scales to obtain multi-scale morphological feature maps; and integrating the multi-scale morphological feature maps to generate the image features.

4. The method according to claim 1, characterized in that, The step of extracting numerical features from the numerical time series data includes: encoding the numerical time series data to obtain data dependency features; extracting structured time fields from the timestamps corresponding to the numerical time series data and encoding them to obtain periodic time features; and fusing the data dependency features and the periodic time features to obtain the numerical features.

5. The method according to claim 1, characterized in that, After fusing the numerical features and the image features, the method further includes: processing the hybrid representation using a cross-representation dual-axis dependency modeling structure to capture dependencies in the representation dimension and the sequence dimension.

6. The method according to claim 5, characterized in that, The cross-representation biaxial dependency modeling structure includes: a representation mixer for transforming along the representation dimension of the mixed representation to model cross-representation interactions; and a sequence mixer for transforming along the sequence dimension of the mixed representation to model temporal dependencies.

7. The method according to claim 1, characterized in that, The scheduling cost-aware loss function includes: an asymmetric quantile loss term, used to apply an asymmetric linear penalty to the prediction error based on a preset ratio of underprediction cost to overprediction cost; and a soft threshold term, used to apply a smooth, simulated one-time penalty when the underprediction error exceeds a preset tolerance band.

8. The method according to claim 7, characterized in that, The soft threshold is implemented using a parameterized sigmoid function, wherein the parameter controls the intensity of the penalty, the tolerance band for triggering the penalty, and the smoothness of the threshold.

9. A time series prediction system, characterized in that, include: The representation conversion module is used to convert the input numerical time-series data into time-series image representations; The feature extraction module is used to extract the numerical features of the numerical time series data and the image features of the time series image representation, respectively. The representation fusion module is used to fuse the numerical features and the image features to obtain a hybrid representation; the prediction and training module is used to generate predicted values ​​based on the hybrid representation and train the model using a scheduling cost-aware loss function, wherein the scheduling cost-aware loss function is used to impose an asymmetric prediction cost on under-prediction and over-prediction.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images