Command and control system resource trend prediction method based on fusion of long and short time sequence characteristics

Abstract

The invention discloses a command and control system resource trend prediction method based on fusion of long and short time sequence characteristics. The method comprises the following steps: acquiring a public power load or similar time sequence monitoring data set, and preprocessing the data in the data set; a deep learning network model based on a TCN-Transformer hybrid model is constructed, a TCN model and a Transformer model are adopted for parallel computing to achieve feature extraction, the TCN model extracts short-term information, the Transformer model extracts long-term features, then fusion features are obtained through a cross attention mechanism and multi-layer perceptron (MLP) weighting, and finally prediction output is generated through full connection layer mapping. Taking data in the training set as input, training the constructed TCN-Transform hybrid model, and continuously optimizing the model until convergence meets a set requirement; and performing prediction by using the trained network model. According to the method, the TCN-Transform hybrid model is constructed, so that local fine-grained features are reserved, the global time trend is effectively captured, and the accuracy of command decision making is improved.

Description

Technical Field

[0001] This invention belongs to the field of time series prediction technology in deep learning, and in particular, it is a method for predicting resource trends in command and control systems based on the fusion of long and short time series features. Background Technology

[0002] Modern command and control systems (such as urban emergency response platforms and intelligent traffic dispatch centers) require real-time coordination of diverse and heterogeneous resources (computing equipment, communication bandwidth, manpower allocation, and energy supply). Dynamic changes in resource status directly impact system response capabilities. Accurately predicting resource demand trends and abnormal evolution paths in advance is a core prerequisite for ensuring the timeliness of command decisions and mitigating systemic risks.

[0003] Existing prediction techniques face key bottlenecks due to the limitations of traditional statistical models and the shortcomings of single deep learning models. For example, methods such as ARIMA exponential smoothing rely on linear assumptions and are difficult to model nonlinear coupling relationships between resource states, such as the surge in computational load leading to increased communication latency. RNN / LSTM suffers from gradient decay in long-sequence scenarios and cannot effectively capture macro-level resource consumption patterns across time periods. TCN, while efficient in extracting local fluctuation features, is not sensitive enough to long-term evolution trends. Transformer excels at modeling global dependencies, but its response to local mutations is lagging, and computational overhead limits real-time performance.

[0004] Existing methods fail to address the deep synergy between short-term fluctuation characteristics and long-term trend characteristics, leading to inaccurate predictions of complex anomalies. In command and control scenarios, resource state evolution exhibits strong spatiotemporal correlation. Current methods suffer from two main shortcomings: they only perform isolated time series predictions; and they lack a cross-scale feature transmission mechanism in the time dimension, making it difficult for short-term anomaly signals to trigger long-term early warnings.

[0005] Therefore, there is an urgent need for a new technical solution to address this problem. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a prediction method based on the TCN-Transformer hybrid model to address the limitations in feature extraction that still exist in existing time series data prediction methods.

[0007] The technical solution of this invention to solve the technical problem is: a method for predicting resource trends in command and control systems based on the fusion of long and short time series characteristics, comprising the following steps, which are performed sequentially:

[0008] Step S1: Obtain publicly available power load or similar time-series monitoring datasets and preprocess the data therein. The datasets include multi-dimensional operating status information at different time points, such as system operating indicators like CPU utilization, memory usage, communication bandwidth load, power consumption, and task scheduling frequency. Subsequently, the processed data is divided into training set, validation set, and test set.

[0009] Step S2: Construct a deep learning network model based on the TCN-Transformer hybrid model. Input the input data processed in step S1 into the TCN model and the Transformer model respectively for parallel feature extraction. Then, obtain the fused features through cross-attention mechanism and multilayer perceptron (MLP) weighting. Finally, generate the trend prediction output of system resources through fully connected layer mapping.

[0010] Step S3: Input the training set data from Step S1 into the deep learning network model based on the TCN-Transformer hybrid model constructed in Step S2 for training. Use mean squared error (MSE) and mean absolute error (MAE) as loss functions. Continuously update the model parameters using the backpropagation algorithm and gradient descent strategy. Monitor the loss changes during the training process and adjust the hyperparameters through the validation set until the model reaches the preset accuracy requirements or convergence conditions on the validation set, thereby obtaining the optimal model parameters.

[0011] Step S4: Use the model trained in step S3 to make predictions on the test set or real-time monitoring data.

[0012] Further, the preprocessing in step S1 includes mean filling for missing values, mean replacement for outliers, and normalization. Missing values ​​refer to unrecorded time-series data points within the acquisition period due to sensor failure, communication delay, or data synchronization interruption. Single missing values ​​are filled with the mean, while continuous missing segments are repaired using a sliding window mean filling method. The window size is dynamically configured based on the resource data type to restore temporal continuity. Outliers refer to outlier data points that deviate from normal variation patterns due to equipment jitter, sudden interference, or transmission errors. An anomaly detection method based on the 3σ principle is used to identify data exceeding the interval [μ-3σ, μ+3σ] as outliers, replacing them with the moving average of adjacent time periods to smooth out abrupt noise. The repaired data is normalized to map features of different dimensions to a unified numerical range. Data timestamp information is parsed, and resampling and alignment are performed based on the minimum sampling interval to eliminate temporal misalignment caused by asynchronous sampling or packet loss.

[0013] Furthermore, the TCN module in step S2 adopts an improved ModernTCN architecture. The time series input first passes through a reversible normalization layer (RevIN) to independently normalize each channel (variable) to eliminate scale differences between different variables. In the prediction stage, the original scale is restored through inverse normalization. Subsequently, the sequence enters the Stem layer, which performs one-dimensional convolution operations using a set kernel size (patch_size) and stride (patch_stride) to divide the long sequence into several time segments (patch) and map them to a higher-dimensional feature space, providing initial feature representations for subsequent temporal modeling. Next, the feature representations sequentially enter multiple Stages (stage modules). Each Stage consists of several Blocks connected in series, where each Block includes two parts: a depthwise separable convolutional (DWConv) module and a reverse bottleneck convolutional feedforward network (ConvFFN) module. As Stages are stacked step by step, the model gradually reduces the resolution of the time dimension through downsampling operations, realizing multi-scale temporal modeling. Finally, after feature fusion and flattened head mapping, the model outputs the target prediction result.

[0014] Furthermore, in step S2, the Transformer module employs a variant iTransformer structure, providing the complete time series for each variable. (Where i represents the variable index and n represents the time step) It is first embedded into a latent representation through a multi-layer perceptron (MLP). This step transforms the original time-series observations into a token-like embedded representation. Its core mechanism models the interdependencies between variables through scaled dot product attention, as shown below:

[0015]

[0016] Where the Q, K, and V matrices are all composed of variables The attention operation is generated by computing attention weights along variable dimension (i); in layer l, the representation of each variable... Independent nonlinear transformations will be performed in the time dimension (n) through a multilayer perceptron-based feed-forward network (FFN):

[0017]

[0018] The final representation of the Lth layer Mapped to predicted values ​​through a multilayer perceptron (MLP) in the output layer. This completes the task of predicting multivariate time series data, as shown in the following formula:

[0019]

[0020] Furthermore, the cross-attention method in step S2 is as follows: the pre-trained TCN module and the Transformer module process the input time series X respectively, generating two different sets of prediction representations:

[0021]

[0022]

[0023] Specifically, the output features of the Transformer module are used to generate the key matrix K and the value matrix V, and the output features of the TCN module are used to generate the query matrix Q. Therefore, the cross-attention calculation is defined as follows:

[0024]

[0025] in , The key / value set and the query set have the same dimension;

[0026] Refined output Projected onto the target output dimension using a multilayer perceptron (MLP), the output is:

[0027]

[0028] in It is a trainable residual scaling factor that dynamically adjusts the interaction weights between attention-enhanced features and the original temporal predictions of the TCN.

[0029] Furthermore, the accuracy determination of model training in step S3 is based on the loss function setting and evaluated using MSE and MAE, the mathematical expressions of which are as follows:

[0030]

[0031]

[0032] in Represents the true value. Represents the predicted value. This represents the number of samples.

[0033] This invention proposes a novel prediction framework: integrating TCN and Transformer models into the same system through a cross-attention fusion mechanism. This design not only preserves local fine-grained features but also effectively captures global temporal trends, while achieving context-aware feature fusion across modules.

[0034] In summary, this invention improves the accuracy of existing time series forecasting methods and can be applied to multiple time series forecasting fields such as command and control. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0036] Figure 1 This is a flowchart of a command and control system resource trend prediction method based on the fusion of long and short time series characteristics, according to the present invention.

[0037] Figure 2 This is an overall structural diagram of the TCN module of the hybrid model of the command and control system resource trend prediction method based on the fusion of long and short time series characteristics, according to the present invention.

[0038] Figure 3 This is a block module structure diagram of the TCN module in the hybrid model of the command and control system resource trend prediction method based on the fusion of long and short time series characteristics, which is a method of the present invention.

[0039] Figure 4 This is an overall structural diagram of the Transformer module of the hybrid model of the command and control system resource trend prediction method based on the fusion of long and short time series characteristics, which is a method of the present invention.

[0040] Figure 5 This is a diagram of the cross-attention module of a hybrid model for a command and control system resource trend prediction method based on the fusion of long and short time series features, as described in this invention.

[0041] Figure 6 This is an overall structural diagram of a hybrid model for a command and control system resource trend prediction method based on the fusion of long and short time series characteristics, as proposed in this invention. Detailed Implementation

[0042] To better understand the present invention, the embodiments of the present invention will be explained in detail below with reference to the accompanying drawings.

[0043] A method for predicting resource trends in command and control systems based on the fusion of long and short time-series characteristics, the overall process framework of which is as follows: Figure 1 As shown, it includes the following steps:

[0044] S1: Obtain publicly available power load or similar time-series monitoring datasets and preprocess the data therein. The datasets include multi-dimensional operating status information at different time points, such as system operating indicators like CPU utilization, memory usage, communication bandwidth load, power consumption, and task scheduling frequency. Subsequently, the processed data is divided into training set, validation set, and test set.

[0045] S2: Construct a deep learning network model based on the TCN-Transformer hybrid model. Input the input data processed in step S1 into the TCN model and the Transformer model respectively for parallel feature extraction. Then, obtain the fused features through cross-attention mechanism and multilayer perceptron (MLP) weighting. Finally, generate the trend prediction output of system resources through fully connected layer mapping.

[0046] S3: Input the training set data from step S1 into the deep learning network model based on the TCN-Transformer hybrid model constructed in step S2 for training. Use mean squared error (MSE) and mean absolute error (MAE) as loss functions. Continuously update the model parameters using the backpropagation algorithm and gradient descent strategy. Monitor the loss changes during the training process and adjust the hyperparameters through the validation set until the model reaches the preset accuracy requirements or convergence conditions on the validation set, thereby obtaining the optimal model parameters.

[0047] S4: Use the model trained in step S3 to make predictions on the test set or real-time monitoring data.

[0048] Further, the preprocessing in step S1 includes mean filling for missing values, mean replacement for outliers, and normalization. Missing values ​​refer to unrecorded time-series data points within the acquisition period due to sensor failure, communication delay, or data synchronization interruption. Single missing values ​​are filled with the mean, while continuous missing segments are repaired using a sliding window mean filling method. The window size is dynamically configured based on the resource data type to restore temporal continuity. Outliers refer to outlier data points that deviate from normal variation patterns due to equipment jitter, sudden interference, or transmission errors. An anomaly detection method based on the 3σ principle is used to identify data exceeding the interval [μ-3σ, μ+3σ] as outliers, replacing them with the moving average of adjacent time periods to smooth out abrupt noise. The repaired data is normalized to map features of different dimensions to a unified numerical range. Data timestamp information is parsed, and resampling and alignment are performed based on the minimum sampling interval to eliminate temporal misalignment caused by asynchronous sampling or packet loss.

[0049] Furthermore, the TCN module in step S2 adopts an improved ModernTCN architecture. The time series input first passes through a reversible normalization layer (RevIN) to independently normalize each channel (variable) to eliminate scale differences between different variables. In the prediction stage, the original scale is restored through inverse normalization. Subsequently, the sequence enters the Stem layer, which performs one-dimensional convolution operations using a set kernel size (patch_size) and stride (patch_stride) to divide the long sequence into several time segments (patch) and map them to a higher-dimensional feature space, providing initial feature representations for subsequent temporal modeling. Next, the feature representations sequentially enter multiple Stages (stage modules). Each Stage consists of several Blocks connected in series, where each Block includes two parts: a depthwise separable convolutional (DWConv) module and a reverse bottleneck convolutional feedforward network (ConvFFN) module. As Stages are stacked step by step, the model gradually reduces the resolution of the time dimension through downsampling operations, realizing multi-scale temporal modeling. Finally, after feature fusion and flattened head mapping, the model outputs the target prediction result.

[0050] Furthermore, in step S2, a Transformer module is used to extract long-term features of the data. The Transformer module employs a variant iTransformer structure, representing the complete time series of each variable. (Where i represents the variable index and n represents the time step) It is first embedded into a latent representation through a multi-layer perceptron (MLP). This step transforms the original time-series observations into a token-like embedded representation. Its core mechanism models the interdependencies between variables through scaled dot product attention, as shown below:

[0051]

[0052] Where the Q, K, and V matrices are all composed of variables The attention operation generates attention weights along variable dimension (i), where, in layer l, the representation of each variable... Independent nonlinear transformations will be performed in the time dimension (n) through a multilayer perceptron-based feed-forward network (FFN):

[0053]

[0054] The final representation of the Lth layer Mapped to predicted values ​​through a multilayer perceptron (MLP) in the output layer. This completes the task of predicting multivariate time series data, as shown in the following formula.

[0055]

[0056] Furthermore, in step S2, to integrate the advantages of high-level feature representations of TCN and Transformer, and effectively capture complementary temporal dynamics and interactions between variables, a cross-attention module is used. The method is as follows: the pre-trained TCN module and the Transformer module process the input time series X respectively, generating two different sets of prediction representations:

[0057]

[0058]

[0059] Specifically, the output features of the Transformer module are used to generate the key matrix K and the value matrix V, and the output features of the TCN module are used to generate the query matrix Q. Therefore, the cross-attention calculation is defined as follows:

[0060]

[0061] in , The key / value set and the query set have the same dimension;

[0062] Refined output The model is projected onto the target output dimension using a multilayer perceptron (MLP). To enhance model stability and expressive power, the final output incorporates residual connections from a temporal convolutional network (TCN) pathway.

[0063]

[0064] in, It is a trainable residual scaling factor that dynamically adjusts the interaction weights between attention-enhanced features and the original temporal predictions of the TCN.

[0065] Furthermore, in step S3, to effectively evaluate the prediction accuracy of the model, the accuracy of model training is determined based on the loss function setting, and evaluated through MSE and MAE, the mathematical expressions of which are as follows:

[0066]

[0067]

[0068] in Represents the true value. Represents the predicted value. This represents the number of samples.

[0069] To facilitate understanding of the present invention, the resource trend prediction method for the command and control system of the present invention will be described below with reference to specific embodiments:

[0070] Example 1

[0071] A method for predicting resource trends in command and control systems based on the fusion of long and short time-series characteristics, comprising the following steps:

[0072] Step S1: Obtain the publicly available power load dataset and preprocess it. The specific steps are as follows:

[0073] Step S1-1 involves preprocessing the acquired power dataset, including filling missing values ​​with the mean, replacing outliers with the mean, and then normalizing. Missing values ​​refer to unrecorded time-series data points within the acquisition period due to sensor failure, communication delay, or data synchronization interruption. Single missing values ​​are filled with the mean, while continuous missing segments are repaired using a sliding window mean filling method. The window size is dynamically configured based on the data type to restore temporal continuity. Outliers refer to outlier data points that deviate from normal variation patterns due to equipment jitter, sudden interference, or transmission errors. An anomaly detection method based on the 3σ principle is used to identify data exceeding the interval [μ-3σ, μ+3σ] as outliers, replacing them with the moving average of adjacent time periods to smooth out abrupt noise. The repaired data is then normalized to map features of different dimensions to a unified numerical range. Data timestamp information is parsed, and resampling and alignment are performed based on the minimum sampling interval to eliminate temporal misalignment caused by asynchronous sampling or packet loss.

[0074] Step S1-2: Divide the dataset into training set, test set, and validation set in a ratio of 7:2:1.

[0075] Step S2: Construct a deep learning network model based on the TCN-Transformer hybrid model. Input the data processed in Step S1 into the TCN model and the Transformer model respectively for parallel feature extraction. Then, obtain the fused features through cross-attention mechanism and multilayer perceptron (MLP) weighting. Finally, generate the trend prediction output of system resources through fully connected layer mapping. The specific steps are as follows:

[0076] Step S2-1: The processed data is fed into the TCN module for feature extraction. The time series input first passes through a reversible normalization layer (RevIN) to independently normalize each channel (variable) to eliminate scale differences between different variables. In the prediction stage, the original scale is restored through inverse normalization. Subsequently, the sequence enters the Stem layer, which performs one-dimensional convolution operations using a set kernel size (patch_size) and stride (patch_stride) to divide the long sequence into several time segments (patch) and map them to a higher-dimensional feature space, providing initial feature representations for subsequent temporal modeling. Next, the feature representations are sequentially fed into multiple Stages (stage modules). Each Stage consists of several Blocks connected in series, where each Block includes two parts: a depthwise separable convolutional (DWConv) module and a reverse bottleneck convolutional feedforward network (ConvFFN) module. As Stages are stacked, the model gradually reduces the resolution of the time dimension through downsampling operations to achieve multi-scale temporal modeling. Finally, after feature fusion and flattened head mapping, the model outputs the target prediction result.

[0077] Step S2-2: The processed data is fed into the Transformer module for feature extraction; the Transformer module uses a variant iTransformer structure, providing the complete time series of each variable. (Where i represents the variable index and n represents the time step) It is first embedded into a latent representation through a multi-layer perceptron (MLP). Then, attention weights are computed along variable dimension (i), where each variable is represented in layer l. It will undergo independent nonlinear transformations in the time dimension (n) through a multilayer perceptron-based feed-forward network (FFN); the final representation of the Lth layer... Mapped to predicted values ​​through a multilayer perceptron (MLP) in the output layer. This allows for the completion of multivariate time series prediction tasks.

[0078] Step S2-3: Extract the prediction features of the pre-trained TCN module and Transformer module respectively. The output features of the Transformer module are used as the key matrix K and the value matrix V, and the output features of the TCN module are used as the query matrix Q. The fused cross-attention feature representation is obtained. The output of the cross-attention mechanism enters the multilayer perceptron (MLP) and then passes through the residual connection layer to obtain the prediction value.

[0079] Step S3: Input the training set data from Step S1 into the deep learning network model based on the TCN-Transformer hybrid model constructed in Step S2 for training. Use mean squared error (MSE) and mean absolute error (MAE) as loss functions. Continuously update the model parameters using the backpropagation algorithm and gradient descent strategy. Monitor the loss changes during the training process and adjust the hyperparameters through the validation set until the model reaches the preset accuracy requirements or convergence conditions on the validation set, thereby obtaining the optimal model parameters.

[0080] Step S4: Use the model trained in step S3 to make predictions on the test set.

[0081] In this embodiment, the constructed dual-channel cross-attention network model was trained with 100 iterations, a lookback window, and a prediction length step of 96. To verify the applicability of the TCN-Transformer hybrid model, it was compared with other models on an electricity dataset, and the results are shown in Table 1. It can be seen that the proposed hybrid cross-attention network model VBTCKN still outperforms other models.

[0082] Table 1 Comparison of Prediction Results

[0083] Model MSE MAE Transformer 0.260 0.358 Reformer 0.312 0.402 Informer 0.274 0.368 Flowformer 0.215 0.320 Flashformer 0.259 0.357 TCN 0.169 0.262 Transformer 0.148 0.240 Combined Prediction Model 0.145 0.238

[0084] Although the specific implementation of the invention has been described above in conjunction with the accompanying drawings and embodiments, it is not intended to limit the scope of protection of the invention. Based on the technical solution of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for predicting resource trends in command and control systems based on the fusion of long and short time-series characteristics, characterized in that: Includes the following steps, And the following steps are performed in sequence; Step S1: Obtain publicly available power load or similar time-series monitoring datasets and preprocess the data. Then, divide the processed data into training set, validation set, and test set. Step S2: Construct a deep learning network model based on the TCN-Transformer hybrid model. Input the input data processed in step S1 into the TCN model and the Transformer model respectively for parallel feature extraction. Then, obtain the fused features through cross-attention mechanism and multilayer perceptron (MLP) weighting. Finally, generate the trend prediction output of system resources through fully connected layer mapping. Step S3: Input the training set data from Step S1 into the deep learning network model based on the TCN-Transformer hybrid model constructed in Step S2 for training. Use mean squared error (MSE) and mean absolute error (MAE) as loss functions. Continuously update the model parameters using the backpropagation algorithm and gradient descent strategy. Monitor the loss changes during the training process and adjust the hyperparameters through the validation set until the model reaches the preset accuracy requirements or convergence conditions on the validation set, thereby obtaining the optimal model parameters. Step S4: Use the model trained in step S3 to make predictions on the test set or real-time monitoring data.

2. The resource trend prediction method for command and control systems based on the fusion of long and short time series characteristics according to claim 1, characterized in that: The dataset mentioned in step S1 includes multi-dimensional operating status information at different time points, such as system operating indicators like CPU utilization, memory usage, communication bandwidth load, power consumption, and task scheduling frequency.

3. The resource trend prediction method for command and control systems based on the fusion of long and short time series characteristics according to claim 1, characterized in that: The preprocessing in step S1 includes filling missing values ​​with the mean, replacing outliers with the mean, and then normalizing. The missing values ​​refer to unrecorded time-series data points caused by sensor failure, communication delay, or data synchronization interruption during the acquisition period. Single missing points are filled with the mean, and continuous missing segments are repaired using the sliding window mean filling method. The window size is dynamically configured according to the data type of the resource to restore the continuity of time. The outliers refer to outlier data points that deviate from the normal variation pattern due to equipment jitter, sudden interference, or transmission errors. An outlier detection method based on the 3σ principle is adopted to identify data points that exceed the interval [μ-3σ, μ+3σ] as outliers and replace them with the moving average of adjacent time periods to smooth out sudden noise. The repaired data is normalized to map features of different dimensions to a unified numerical range. Parse the data timestamp information, and resample and align it based on the minimum sampling interval to eliminate timing misalignment caused by asynchronous sampling or packet loss.

4. The resource trend prediction method for command and control systems based on the fusion of long and short time series characteristics according to claim 1, characterized in that: The TCN module in step S2 adopts the improved ModernTCN architecture. The time series input first passes through the reversible normalization layer (RevIN) to independently normalize each channel variable in order to eliminate the scale difference between different variables. In the prediction phase, the original scale is restored through inverse normalization; Subsequently, the sequence enters the Stem layer, which performs one-dimensional convolution operations using a set kernel size (patch_size) and stride (patch_stride) to divide the long sequence into several time segments (patch) and map them to a higher-dimensional feature space, providing initial feature representations for subsequent temporal modeling. Next, the feature representations are sequentially fed into multiple Stage modules. Each Stage consists of several Blocks connected in series, and each Block includes two parts: a Depthwise Separable Convolutional Network (DWConv) module and a Reverse Bottleneck Convolutional Feedforward Network (ConvFFN) module. As Stages are stacked, the model gradually reduces the resolution of the temporal dimension through downsampling operations, achieving multi-scale temporal modeling. Finally, after feature fusion and flattened head mapping, the model outputs the target prediction result.

5. The resource trend prediction method for command and control systems based on the fusion of long and short time series characteristics according to claim 1, characterized in that: In step S2, the Transformer module employs a variant iTransformer structure, providing a complete time series for each variable. Where i represents the variable index and n represents the time step, it is first embedded into a latent representation through a multilayer perceptron (MLP). It transforms the original time-series observations into a token-like embedded representation. Its core mechanism models the interdependencies between variables through scaled dot product attention, as shown below: The Q, K, and V matrices are all composed of variables. The attention operation computes attention weights along the variable dimension (i), thereby enabling explicit modeling of cross-variable correlations. In the l-th layer, the representation of each variable... It will perform independent nonlinear transformations in the time dimension (n) through a multilayer perceptron-based feedforward network FFN: The final representation of the Lth layer The output layer is mapped to the predicted value by a multilayer perceptron (MLP). This completes the task of predicting multivariate time series data, as shown in the following formula: 。 6. The resource trend prediction method for command and control systems based on the fusion of long and short time series characteristics as described in claim 1, characterized in that: The cross-attention method in step S2 is as follows: the pre-trained TCN module and the Transformer module process the input time series X respectively, generating two different sets of prediction representations: The output features of the Transformer module are used to generate the key matrix K and the value matrix V, and the output features of the TCN module are used to generate the query matrix Q. Therefore, the calculation of cross-attention is defined as follows: in , The key / value set and the query set have the same dimension; Refined output Projected onto the target output dimension by a multilayer perceptron (MLP), the final output is: in It is a trainable residual scaling factor that dynamically adjusts the interaction weights between attention-enhanced features and the original temporal predictions of the TCN.

7. The resource trend prediction method for command and control systems based on the fusion of long and short time series characteristics according to claim 1, characterized in that: In step S3, the accuracy of model training is judged based on the loss function, and evaluated using MSE and MAE. The mathematical expressions are as follows: in Represents the true value. Represents the predicted value. This represents the number of samples.

Images