A time point value and interval value synchronous prediction method based on a double-decoding generation network

Abstract

The application discloses a kind of time series point value and interval value synchronous prediction method based on double decoding generation network, belong to time series prediction technical field.For existing time series prediction method is difficult to output point value and real physical interval value simultaneously, probability type interval lacks physical meaning, interval prediction relies on sliding window and leads to low reliability and other problems, the application constructs the double decoding parallel architecture of the unification of space-time feature encoder, point prediction decoder and interval prediction decoder three;Global overall approximation strategy is adopted, and the interval distribution of the whole time series is directly overall fitted, and the upper and lower limits of the complete interval are output at one time;Through multiple composite loss function, the interval generation is finely constrained from five dimensions of boundary effectiveness, matching degree, coverage, width rationality and reconstruction error.The model is trained in stages, and the point value and interval value prediction results can be output simultaneously after one forward inference.The application is verified on measured data set, and the point prediction accuracy and interval reliability are significantly better than existing models, which can adapt to real-time collaborative prediction scenarios such as environmental monitoring, financial analysis, unmanned system and other scenarios that need to obtain point value and interval value simultaneously.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and in particular to a prediction method for multivariate time series data that enables the synchronous output of point values ​​and real physical interval values. Background Technology

[0002] Against the backdrop of the rapid development of data intelligence and the Industrial Internet of Things, time series data has become a core data carrier in fields such as environmental monitoring, financial quantitative analysis, and intelligent unmanned systems. Time series forecasting technology, by mining the time dependencies, trend patterns, and nonlinear variation modes in historical data, enables numerical inferences about future states, and is a key foundation for achieving in-depth analysis, autonomous perception, and intelligent decision-making in various systems.

[0003] In many practical applications, the observed time-series data is not in a single form, but rather contains both point values ​​and interval values. Time-series point values ​​are single observations that change over time, such as real-time temperature in meteorological monitoring, stock closing prices in financial trading, position coordinates in unmanned surface vessel / vehicle control systems, and measurements from industrial sensors. Time-series interval values ​​are bandwidth-type data composed of the upper and lower bounds of real physical quantities, such as daily high / low temperatures, intraday high / low prices of stocks, upper / lower limits of sensor noise fluctuations, and the offset range of pose data due to noise interference. This type of interval data has a clear physical meaning and is not a virtual probability interval derived from point prediction.

[0004] Most existing time series prediction techniques only focus on point values, and the predicted interval values ​​mainly refer to the fluctuation probability of point values, rather than the upper and lower bounds of the actual physical quantities. Existing techniques mostly adopt a model of predicting point values ​​first and then constructing probability intervals, which can be mainly divided into the following three categories: 1) Quantile regression method, which takes the point value prediction result as the center and generates probability intervals with different confidence levels by performing regression modeling on different quantiles; 2) Variance estimation and probability distribution method, which uses the point value prediction result as the mean and constructs a probability distribution by estimating the variance of the prediction error, thereby generating uncertainty intervals; 3) Deep learning estimation method, which outputs probability intervals while outputting point predictions by modifying the network structure or loss function.

[0005] The above methods all treat interval values ​​as a byproduct of time series point value prediction, which has obvious defects: 1) The intervals do not have real physical meaning. The intervals generated by the existing methods are virtual probability ranges, rather than the upper and lower bounds of real observations, which cannot match the actual engineering needs; 2) The coordination between point values ​​and interval values ​​is poor. Step-by-step modeling leads to inconsistencies between point and interval results, and they cannot be output synchronously in a forward inference; 3) The loss function has a single constraint dimension, focusing only on reconstruction error and lacking multi-dimensional constraints on boundary validity, interval coverage, and width rationality; 4) The multivariate spatiotemporal correlation mining is insufficient. It does not fully integrate the spatial correlation between variables and the long-term dependence of the time dimension, resulting in low feature utilization.

[0006] Current prediction techniques primarily target probabilistic intervals, such as confidence intervals and uncertainty quantification intervals, rather than data intervals defined by the upper and lower bounds of actual physical observations. While these methods provide a measure of prediction uncertainty, they cannot reconstruct the true physical intervals in actual observations. How to directly perform end-to-end predictions on observation interval data with real physical significance remains a pressing technical problem in this field. Furthermore, existing technologies cannot achieve direct, synchronous, and high-precision joint prediction of time-series point data and true physical interval data, making it difficult to meet the practical requirements of scenarios such as environmental monitoring, financial analysis, and unmanned system control. Therefore, this paper proposes a novel time-series analysis method capable of simultaneously predicting point values ​​and interval values, which has significant theoretical and engineering application value. Summary of the Invention

[0007] To address the problems in existing technologies, such as the difficulty in synchronously predicting time series point values ​​and real physical interval values, the lack of real physical meaning in interval values, insufficient continuity and reliability due to the reliance on sliding windows for interval prediction, and the single dimension of the loss function for interval constraints, this invention provides a method for synchronously predicting time series point values ​​and interval values ​​based on a dual-decoding generation network.

[0008] Unlike existing technologies, this invention has the following three core innovative features: First, it proposes a dual-decoding parallel architecture, treating point prediction and real physical interval prediction as two parallel and independent decoding tasks for the first time. It constructs a unified model architecture of spatiotemporal feature encoder, point prediction decoder, and interval prediction decoder, achieving simultaneous output of point prediction values ​​and interval prediction values ​​in a single forward inference. Second, it proposes a global approximation interval generation strategy. Unlike the sliding window local prediction method commonly used in existing interval prediction, this invention employs a global approximation strategy, directly fitting the interval distribution of the entire time series to output complete upper and lower limits of the intervals in one go, ensuring the continuity and global consistency of interval predictions. Third, it proposes a multi-dimensional constraint mechanism with multiple composite losses. It constructs a loss function from five dimensions: interval boundary validity, prediction matching degree, real value coverage, interval width rationality, and overall reconstruction error, achieving comprehensive and refined constraints on the interval generation process.

[0009] To achieve the above-mentioned objectives, the present invention is implemented through the following technical solution: A method for synchronous prediction of time series point values ​​and interval values ​​based on a dual-decoding generator network includes the following steps:

[0010] A spatiotemporal feature encoder, a point prediction decoder, and an interval prediction decoder are built sequentially to form a dual-decoding generation network architecture with shared encoders and parallel collaboration between the two decoders.

[0011] The spatiotemporal feature encoder consists of a spatial feature extraction module, a temporal feature extraction module, and an adaptive attention fusion module. The spatial feature extraction module constructs a fully connected adjacency matrix based on the correlations between variables in multivariate time-series data and performs self-loop normalization on it. It then extracts spatial correlation features between variables using a graph convolutional network. The temporal feature extraction module captures the long-term dependencies and evolutionary patterns of time-series data using gated recurrent units, generating temporal dependency features. The adaptive attention fusion module weights and fuses the spatial correlation features and temporal dependency features, adaptively assigning feature weights to generate a global spatiotemporal feature matrix that considers both the spatial correlations of variables and their dynamic temporal evolution, serving as the input features shared by subsequent decoders.

[0012] The point prediction decoder and the interval prediction decoder adopt a parallel structure and share the global spatiotemporal feature matrix.

[0013] The point prediction decoder is built upon a conditional generative adversarial network (GAN) and incorporates residual connections to enhance feature propagation efficiency. The decoder comprises a point prediction generator and a point prediction discriminator. The point prediction generator takes global spatiotemporal features, random noise, and historical time-series windows as conditional inputs to generate time-series point predictions; the point prediction discriminator distinguishes between the generated point predictions and the actual point values. The point prediction decoder is trained using a joint optimization of adversarial loss and mean squared error reconstruction loss to continuously improve point prediction accuracy.

[0014] The interval prediction decoder is built on a generative adversarial network architecture and employs a global approximation strategy to achieve interval prediction. The decoder consists of an interval prediction generator and an interval prediction discriminator. The interval prediction generator abandons the traditional sliding window local prediction mode, taking global spatiotemporal features and random noise as input, and directly performs a global fit on the interval distribution of the entire time series, outputting the upper and lower bound sequences of the complete prediction interval in one go. The interval prediction discriminator distinguishes between the generated prediction interval and the actual physical interval.

[0015] Data cleaning was performed on the original multivariate time-series data, including outlier removal and missing value imputation, to ensure the continuity and integrity of the time-series data. Max-min standardization was then applied to the cleaned time-series data to eliminate differences in units and numerical scales between different variables. Simultaneously, a real physical interval dataset corresponding to the point data was constructed based on actual observation data, forming a complete dataset containing the time-series point sequences and their corresponding real upper and lower bound intervals.

[0016] The standardized time-series data is input into the spatiotemporal feature encoder to extract spatial correlation features and temporal dependency features respectively. After attention fusion, the global spatiotemporal feature matrix is ​​obtained.

[0017] The network model training employs a phased training strategy. First, the spatiotemporal feature encoder and point prediction decoder are trained jointly to optimize feature extraction capabilities and point prediction accuracy. Second, the parameters of the spatiotemporal feature encoder are fixed, and the interval prediction decoder is trained separately to improve the reliability of interval prediction. Finally, all parameter constraints are removed, and the spatiotemporal feature encoder, point prediction decoder, and interval prediction decoder are treated as a whole for end-to-end collaborative fine-tuning. This eliminates inconsistencies between point prediction and interval prediction results, ensuring that the predicted point value is physically within the prediction interval.

[0018] During the training of the interval prediction decoder, a multi-factor composite loss function is employed for optimization. This loss function integrates the following five constraints: 1) interval boundary validity constraint, ensuring that the lower bound of the predicted interval is less than or equal to the upper bound; 2) matching degree constraint between the predicted interval and the true interval, making the predicted interval approximate the true interval in shape and position; 3) true value coverage constraint, ensuring that the true observation point value falls within the predicted interval with a high probability; 4) reasonable interval width constraint, preventing the predicted interval from being too wide, resulting in reduced information, or too narrow, resulting in insufficient coverage; 5) overall interval reconstruction error constraint, ensuring that the generated interval is close to the true interval in terms of data distribution. Through the weighted joint optimization of the above multi-factor composite loss function, refined constraints and training of the interval prediction decoder are achieved.

[0019] After the network model is trained, the multivariate time series data to be predicted is input into the trained model. The model automatically normalizes the data to be predicted, and outputs point prediction results and interval prediction results simultaneously through one forward inference. After destandardization, the original dimensions of the data are restored to obtain the final time series point prediction values ​​and interval prediction values.

[0020] The model's predictive performance was evaluated across two task dimensions: point prediction and interval prediction. Point prediction accuracy was assessed using metrics such as root mean square error, mean absolute error, and coefficient of determination, comprehensively measuring the deviation between predicted and actual values ​​and the goodness of fit. Interval prediction reliability was evaluated using the interval overlap rate, quantifying the consistency between the predicted and actual intervals in spatial distribution. Through this metric system, the model's overall performance in terms of point prediction accuracy, data fit, and interval prediction matching was comprehensively verified. Beneficial effects

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. By constructing a unified parallel architecture of spatiotemporal feature encoder, point prediction decoder and interval prediction decoder, end-to-end synchronous prediction of temporal point values ​​and real physical interval values ​​is achieved for the first time. This overcomes the inconsistency problem of point and interval results caused by traditional step-by-step modeling with points first and then intervals. Complete prediction information can be obtained with a single forward inference, which significantly improves prediction efficiency and synergy.

[0022] 2. A global approximation strategy is adopted to replace the traditional sliding window local prediction method. The entire time series interval distribution is directly modeled and fitted as a whole, which fundamentally ensures the continuity and global consistency of the prediction interval in the time dimension and avoids the problems of interval fragmentation and boundary mismatch caused by local prediction.

[0023] 3. A multi-composite loss function was designed to comprehensively constrain and finely optimize the interval generation process from five dimensions: boundary validity, matching degree, coverage, width rationality, and reconstruction error. This ensures that the generated prediction intervals not only have real physical meaning, but also outperform existing methods in terms of accuracy, reliability, and practicality.

[0024] 4. The spatiotemporal feature encoder, through deep fusion of graph convolutional network and gated recurrent unit and the introduction of adaptive attention mechanism, can simultaneously capture the spatial correlation and temporal dynamic evolution law in multivariate time series data, providing high-quality shared feature representation for subsequent dual decoders and effectively improving the model's ability to model complex spatiotemporal dependencies.

[0025] 5. The phased training strategy effectively balances the optimization conflict between point prediction and interval prediction during the training process. The approach of first optimizing independently and then fine-tuning collaboratively ensures the stability and convergence of the model training, so that the final model achieves the best level in both point prediction accuracy and interval reliability. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Attached Figure Description

[0027] Figure 1 The flowchart shows the model training process of the synchronous prediction method for time series point values ​​and interval values ​​based on a dual-decoding generation network provided in this embodiment of the invention.

[0028] Figure 2 This is a schematic diagram of the dual-decoding generation network provided in an embodiment of the present invention.

[0029] This invention provides a method for simultaneous prediction of temporal point values ​​and interval values ​​based on a dual-decoding generator network. The overall training process is as follows: Figure 1 As shown, the network structure is as follows Figure 2 As shown in the attached diagram, this method comprises three main steps: network construction, data preprocessing and model training, and model application and evaluation. The specific implementation of each step is explained in detail below with reference to the accompanying diagram.

[0030] This step sequentially constructs a spatiotemporal feature encoder, a point prediction decoder, and an interval prediction decoder, forming a three-in-one dual-decoding generative network architecture with shared encoders and parallel collaboration between the two decoders. Figure 2 As shown.

[0031] The first step involves constructing a spatiotemporal encoder, which consists of a spatial feature extraction module, a temporal feature extraction module, and an adaptive attention fusion module. This module generates a global spatiotemporal feature matrix that takes into account both the spatial correlation of variables and the dynamic evolution of time, serving as the shared input features for subsequent point prediction decoders and interval prediction decoders.

[0032] Suppose the multivariate time series after Min-Max standardization is:

[0033] in, This represents the input multivariate time series. Represents the multivariable state vector at time t; Let C represent the sequence length and C represent the number of variables. First, let... Multivariate time-series data is input into a fully connected critical matrix constructed based on the correlation of individual state variables. This matrix is ​​used to avoid spatial correlation between variables. The adjacency matrix is ​​defined as follows:

[0034] in It is an adjacency matrix. is the dimension of the adjacency matrix, which is consistent with the number of variables.

[0035] To avoid gradient vanishing during graph convolution, we perform a normalization process with self-loops on the adjacency matrix. The normalization formula is as follows:

[0036] in The normalized adjacency matrix, It is the identity matrix. Let be the degree matrix of the adjacency matrix. It is a diagonal matrix, representing the connectivity of each node.

[0037] Then, the state vector at each time step is input into the multi-layer GCN to extract the spatial features between variables. The forward propagation formula for a single time step GCN is:

[0038] in Indicates the first The spatial feature vector output by the layer GCN at time step 1; This indicates the number of layers in the GCN, and ReLU represents the activation function. It indicates that GCN is in the The learnable weight matrix of the layer.

[0039] Finally, after passing through GCN, the resulting spatial feature matrix is:

[0040] in This represents the output spatial feature sequence of GCN. This represents the spatial feature vector extracted by multiple GCNs at time step i. This represents the spatial feature dimension of the GCN output.

[0041] The GRU temporal feature extraction layer is used to capture the evolutionary patterns and long-term dependencies of time-series data. The forward propagation formula of GRU is: Where GRU stands for Gated Cyclic Unit. This represents the feature sequences that GRU does not output at any time. Represents the hidden state of the last time step in the GRU. Then, the output features of the GRU are dimension-mapped through a fully connected layer to obtain a temporal feature sequence that matches the spatial orthogonal dimension. The formula is as follows:

[0042] in This represents the mapped temporal feature sequence. Indicates a fully connected layer. This represents the dimension of the GRU output, compared to that in GCN. The output dimensions are consistent.

[0043] The attention fusion layer concatenates the spatial features output by the GCN with the temporal features output by the GRU along the channel dimension to obtain a spatiotemporal feature sequence, the formula of which is:

[0044] in Indicates the first The spatiotemporal feature vector at time t, The dimension representing the joint feature is the sum of the spatial feature dimension and the temporal feature dimension.

[0045] Then, the attention weights at each time step are calculated using two fully connected layers to adaptively allocate the importance of different features at different time steps. The formula is as follows:

[0046] in Let be the attention weight at time i. and These represent two fully connected layers for attention, and tanh represents the activation function. This represents the normalization function.

[0047] The formula for the spatiotemporal features after the above attention mechanism fusion process is:

[0048] The second step involves constructing a point prediction decoder, which is based on a conditional generative adversarial network and incorporates residual connections to enhance feature propagation efficiency. Its structure is as follows: Figure 2 The midpoint prediction decoder section is shown. This decoder contains a point prediction generator and a point prediction discriminator.

[0049] The point prediction generator is a conditional generator that takes random noise, historical time-series window features, and global spatiotemporal features as conditional inputs. It improves feature propagation efficiency through residual connections and outputs predicted point values, with the following formula:

[0050] in For point prediction generator, For point prediction results, It is a random noise vector. The spatiotemporal characteristic matrix, As a historical timeline window, For batch size, This represents the number of variables to be predicted.

[0051] The point prediction discriminator is a conditional discriminator that takes real point data and historical time series windows as inputs, determines the authenticity of the data generated by the generator, and outputs the discrimination probability. Its formula is:

[0052] in Indicates a point prediction discriminator. This is the actual value.

[0053] The point prediction decoder employs joint optimization of adversarial loss and reconstruction loss to ensure the accuracy of the generated point predictions. Its formula is as follows:

[0054] in The total loss of the point prediction conditional discriminator, For mathematical expectation, For the distribution of real points, Given a random noise distribution, log represents the logarithmic function used to calculate the cross-entropy loss. The discriminant loss represents the true data. Discriminative loss in generated data.

[0055] The generator loss function, used to optimize the accuracy and realism of point predictions generated by the generator, is formulated as follows:

[0056] in This represents the total loss of the point prediction condition generator. This represents the adversarial loss of the generator. This represents the reconstruction loss weighting coefficient. This represents the MSE reconstruction loss.

[0057] The third step involves constructing the interval prediction decoder, which is based on a generative adversarial network architecture and employs a global approximation strategy to achieve interval prediction. Its structure is as follows: Figure 2 The section on interval prediction decoder is shown. This decoder includes an interval prediction generator and an interval prediction discriminator.

[0058] The interval prediction generator takes the spatiotemporal feature matrix output by the encoder and random noise as input, and calculates the output interval prediction result using the following formula:

[0059] in For interval prediction generator, For interval prediction results, For interval noise vectors, This is the nth prediction interval.

[0060] The interval prediction discriminator is used to determine the authenticity of the prediction intervals generated by the generator. Its formula is:

[0061] in This represents an interval prediction discriminator. Represents the true interval data First, there is the interval boundary loss, which is used to ensure the physical validity of the interval. Its formula is:

[0062] in For interval boundary constraint loss, For boundary margin Secondly, there is the boundary matching loss, which ensures that the predicted interval covers the true interval. Its formula is:

[0063] in For boundary matching loss, For the prediction interval, This represents the true interval.

[0064] Next is the interval coverage loss, which is a complex loss function that ensures the predicted interval contains the true value. Its formula is:

[0065] in For interval coverage loss, Let be the true state of the i-th variable.

[0066] Next is the interval width penalty, used to avoid intervals that are too wide or too narrow, ensuring the reasonableness of the intervals. Its formula is:

[0067] Among them Interval width penalty loss, Let be the width of the i-th interval. Let be the average width of the th real interval. Next is the interval reconstruction loss, which ensures that the generated interval distribution closely approximates the true interval. Its formula is:

[0068] Finally, the total loss of the interval generator is calculated by weighting the above loss functions to achieve multi-dimensional optimization. The formula is as follows:

[0069] in This represents the weight parameters for each loss.

[0070] First, data preprocessing is performed to clean the original multivariate time series data, including outlier removal and missing value imputation, to ensure the continuity and integrity of the time series data. Specifically, outliers can be identified and removed using the 3σ criterion based on statistical distribution, and missing values ​​can be imputed using linear interpolation or mean imputation methods. After cleaning, the time series data undergoes min-max standardization, mapping the values ​​of each variable to the [0,1] interval to eliminate differences in dimensionality and numerical scale between different variables. The formula is as follows: Secondly, the model training uses, as follows: Figure 1 The phased training strategy shown includes the following three phases: Phase 1: Joint Training of the Spatiotemporal Feature Encoder and Point Prediction Decoder. Standardized time-series point data is input into the spatiotemporal feature encoder, which extracts global spatiotemporal features according to the process described in Step 1 of Phase 1. These features, along with random noise and historical time-series windows, are input into the point prediction generator, where a point prediction discriminator distinguishes the generated results from the true point values. This phase jointly optimizes the parameters of both the spatiotemporal feature encoder and the point prediction decoder, aiming to minimize the total loss of the point prediction decoder, with a focus on optimizing feature extraction capability and point prediction accuracy. During training, the parameters of both the spatiotemporal feature encoder and the point prediction generator are updated simultaneously. Through the dual constraints of adversarial training and minimizing reconstruction error, the encoder learns the feature representation most advantageous for the point prediction task.

[0071] Phase Two: Training a Fixed Spatiotemporal Feature Encoder and a Separate Interval Prediction Decoder. After training in Phase One, the parameters of the fixed spatiotemporal feature encoder remain unchanged. Global spatiotemporal features and random noise are input into the interval prediction generator to generate a complete sequence of predicted intervals; the interval prediction discriminator distinguishes between the generated intervals and the true intervals. This phase optimizes only the parameters of the interval prediction decoder to minimize its multiple composite losses.

[0072] Phase 3: Overall Co-tuning. After the second phase of training, all parameter constraints are removed, and the spatiotemporal feature encoder, point prediction decoder, and interval prediction decoder are treated as a whole for end-to-end co-tuning. The optimization objective of this phase is the joint loss function:

[0073] By optimizing end-to-end collaboration, the inconsistency between point prediction and interval prediction results is further eliminated, ensuring that the point prediction value is physically within the prediction interval, and improving the collaboration of the dual decoder outputs.

[0074] The rationale for adopting the aforementioned phased training approach is as follows: Although point prediction and interval prediction share the same spatiotemporal feature encoder, their optimization objectives are fundamentally different—point prediction focuses on minimizing the accuracy of a single point, while interval prediction focuses on the consistency and reliability of the overall distribution. If end-to-end joint optimization is performed in the early stages of training, the backpropagation of gradients with two different properties can easily lead to gradient conflicts and training instability. By first independently optimizing the point prediction task to establish a stable feature space, then training the interval prediction task based on this feature space, and finally performing collaborative fine-tuning, the optimization conflicts between the two tasks can be effectively balanced, ensuring the stability and convergence of the model training.

[0075] After model training is complete, the multivariate time-series data to be predicted is input into the trained model. The model automatically performs the same max-min standardization process on the data to be predicted as during the training phase. The standardized data is then processed by a spatiotemporal feature encoder to extract global spatiotemporal features, which are subsequently input into the point prediction decoder and the interval prediction decoder, respectively. The point prediction generator outputs point prediction values, and the interval prediction generator outputs interval prediction values. The outputs of both decoders are then de-standardized to restore the original dimensions of the data.

[0076] The model's predictive performance was evaluated for both point prediction and interval prediction tasks. Point prediction accuracy was evaluated using three metrics: root mean square error, mean absolute error, and coefficient of determination.

[0077]

[0078]

[0079]

[0080] in For the true value, For predicted values, This is the upper bound of the predicted value. This is the lower bound of the predicted value. The upper bound of the true value, This is the lower bound of the true value. Example

[0081] To verify the technical effectiveness of the method of the present invention in practical application scenarios, an experiment is conducted using a multivariate time series dataset as an example.

[0082] The dataset used contains 15 variables, covering five time-series point values: Roll, Pitch, Yaw, X, and Y, as well as ten upper and lower bounds for each variable affected by noise. Each variable in the dataset has 1517 rows, with the first 1300 rows for training and 1300-1400 rows for the test set. Regarding model settings, the sliding window size for both the temporal feature encoder and the point prediction decoder is set to 10 steps with a step size of 1; the interval prediction decoder, due to its global approximation approach, does not have a sliding window.

[0083] To reasonably verify the advantages of this invention, we selected typical time-series data prediction models such as RNN, GRU, BESN, Transformer, Informer, and GAN as benchmarks. We employed a multi-input multi-output prediction method to simultaneously predict 15 variables in the dataset. To ensure the fairness and reasonableness of the comparative experiments, the hyperparameters of each comparison model were optimized using a grid search method to determine the relatively optimal hyperparameter combination for this experimental dataset, minimizing the impact of hyperparameter settings on the model performance comparison results. Specific experimental results are shown in Table 1. Table 1 Evaluation Parameters of Comparative Experimental Models

[0084] As can be seen from the evaluation results in Table 1, the proposed dual-decoding generative network model achieves the best performance in all evaluation metrics for point prediction and interval prediction, demonstrating a significant performance advantage over other comparative models. Regarding point prediction accuracy, the dual-decoding generative network model has an average RMSE of 12.558, the lowest among all models, reducing it by approximately 23.9% compared to the second-best Informer model (16.507), and by 56.3%, 61.1%, and 63.6% respectively compared to traditional models such as GRU (28.723), BESN (32.307), and RNN (34.472). The dual-decoding generative network model also has an average MAE of 9.704662, the best among all models, reducing it by 23.7% compared to the Informer model (12.718) and by 66.6% compared to the RNN model (29.095), significantly lower than other comparative models. This fully demonstrates that the dual-decoding generative network model has the smallest point prediction bias and significantly better accuracy in predicting the true values ​​of the data. In terms of data fitting ability, the dual-decode generator network model achieved an average R² of 0.920296, which is 2.86% and 2.62% higher than the two mainstream time series prediction models, Informer (0.894) and Transformer (0.896), respectively. Compared with models such as BESN (0.662) and RNN (0.721), the improvement is over 30%, indicating that the dual-decode generator network model has the best fitting effect on the inherent changing trends and nonlinear laws of time series data, and can more accurately capture the overall evolution characteristics of the data. In the univariate dimension, its pitch angle R² is slightly higher than that of Informer. The R² of displacement relative to the X-axis is basically the same as that of Transformer and is the best of all models. The R² of displacement relative to the Y-axis is far superior to models with poor fitting effects such as RNN and BESN, achieving a high degree of fitting for all variables, which fully demonstrates its best ability to capture the inherent changing trends and nonlinear correlations of time series data.

[0085] In interval prediction, the dual-decode generator network model exhibits an absolute advantage, with an interval overlap rate of 61.23%, far exceeding all comparable models. This value is 2.75 times that of the second-best performing Informer model (22.26%), 3.16 times that of the Transformer model (19.35%), more than 4 times that of GAN (13.90%), BESN (15.12%), and RNN (19.35%), and 6.22 times that of the GRU model (9.84%). The interval overlap rates of the remaining comparable models are all below 23%, with GRU, BESN, and GAN models having interval overlap rates of less than 16%. The experimental results show that the average interval overlap rate (IOR) of the dual-decode generator network reaches 61.23%, far exceeding all other models, indicating that the reliability of this invention in interval prediction is far superior to existing methods. The experimental results above fully demonstrate that the dual-decoding generator network achieves collaborative optimization of point and interval data. In scenarios requiring simultaneous prediction of both point and interval data, its overall prediction performance significantly outperforms existing mainstream time series prediction models. The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for synchronous prediction of time-series point values ​​and interval values ​​based on a dual-decoding generator network, characterized in that, Includes the following steps: Step 1, Network Model Construction: The spatiotemporal feature encoder, point prediction decoder, and interval prediction decoder are built sequentially to form a dual-decoding generation network architecture with shared encoder and parallel collaboration of the two decoders; the spatiotemporal feature encoder is used to extract the global spatiotemporal feature matrix of the input multivariate time series data, which serves as the shared input feature of the point prediction decoder and the interval prediction decoder. The point prediction decoder is used to generate time series point prediction values, and the interval prediction decoder adopts a global approximation strategy to directly fit the interval distribution of the entire time series and output the complete upper and lower limit sequences of the interval at one time. Step 2, Data Preprocessing and Model Training: The original multivariate time series data is cleaned and standardized to construct a real physical interval dataset corresponding to the point data; the processed data is then input into the spatiotemporal feature encoder to extract the global spatiotemporal feature matrix. A phased training strategy is adopted to sequentially complete the joint training of the spatiotemporal feature encoder and the point prediction decoder, the separate training of the interval prediction decoder with a fixed encoder, and the overall collaborative fine-tuning of the encoder and the dual decoders. Among them, the training of the interval prediction decoder adopts multiple composite loss functions to constrain and optimize the interval generation process from five dimensions: interval boundary validity, prediction matching degree, true value coverage, interval width rationality, and overall reconstruction error. Step 3, Model Application and Evaluation: Input the time series data to be predicted into the trained model, and output the point prediction results and interval prediction results simultaneously through one forward inference. The original dimensions are restored by destandardization. The model performance is evaluated by the point prediction accuracy index and the interval prediction reliability index respectively.

2. The method according to claim 1, characterized in that, The spatiotemporal feature encoder described in step 1 consists of a spatial feature extraction module, a temporal feature extraction module, and an adaptive attention fusion module. The spatial feature extraction module constructs a fully connected adjacency matrix based on the correlation between multiple variables, and after normalization with self-loops, extracts spatial correlation features between variables through a graph convolutional network. The temporal feature extraction module captures the long-term dependencies and evolutionary patterns of time-series data in the time dimension through gated recurrent units, generating time-dependent features. The adaptive attention fusion module performs weighted fusion of spatial correlation features and time-dependent features, adaptively assigns feature weights, and generates the global spatiotemporal feature matrix.

3. The method according to claim 1, characterized in that, The point prediction decoder described in step 1 is constructed based on a conditional generative adversarial network and incorporates residual connections. It includes a point prediction generator and a point prediction discriminator. The point prediction generator takes global spatiotemporal features, random noise, and historical time series windows as conditional inputs and outputs time series point prediction values. The point prediction discriminator distinguishes between real and generated point data; the point prediction decoder is trained using a joint optimization of adversarial loss and mean square error reconstruction loss.

4. The method according to claim 1, characterized in that, The interval prediction decoder described in step 1 is built on a generative adversarial network architecture and includes an interval prediction generator and an interval prediction discriminator. The interval prediction generator takes global spatiotemporal features and random noise as input, does not rely on a sliding window, and directly fits the interval distribution of the entire time series to the whole, outputting the upper and lower limits of the intervals for all prediction times at once. The interval prediction discriminator distinguishes between the real physical intervals and the generated intervals.

5. The method according to claim 4, characterized in that, The multiple composite loss function includes: (1) Interval boundary validity constraint loss, used to ensure that the lower bound of the prediction interval is less than or equal to the upper bound; (2) The matching degree constraint loss between the predicted interval and the true interval is used to make the predicted interval approximate the true interval in shape and position; (3) True value coverage constraint loss, used to ensure that the true observation point value falls within the prediction interval with a high probability; (4) Interval width reasonableness constraint loss, used to prevent the prediction interval from being too wide or too narrow; (5) Interval overall reconstruction error constraint loss, used to ensure that the generated interval is close to the real interval in terms of data distribution; The five loss terms mentioned above, after being weighted and combined, together with the adversarial loss, constitute the total loss function of the interval prediction generator.

6. The method according to claim 1, characterized in that, The data preprocessing described in step 2 specifically includes: removing outliers and filling missing values ​​in the original multivariate time series data; performing max-min standardization on the time series point data; constructing real physical interval data corresponding to the point data based on the actual observation data; and forming a complete dataset containing the time series point sequence and the real upper and lower bound intervals.

7. The method according to claim 1, characterized in that, The phased training strategy described in step 2 specifically includes: In the first stage, the spatiotemporal feature encoder and the point prediction decoder are trained together to optimize feature extraction capability and point prediction accuracy. In the second stage, the parameters of the spatiotemporal feature encoder are fixed, and the interval prediction decoder is trained separately to improve the reliability of interval prediction. In the third stage, all parameter restrictions are removed, and the spatiotemporal feature encoder, point prediction decoder, and interval prediction decoder are treated as a whole for end-to-end collaborative fine-tuning to eliminate inconsistencies between point prediction and interval prediction results.

8. The method according to claim 1, characterized in that, The point prediction accuracy indicators mentioned in step 3 include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The interval prediction reliability indicator is interval overlap rate (IOR).

9. The method according to any one of claims 1 to 8, characterized in that, The method is applied to multivariate time series data prediction scenarios in environmental monitoring, financial analysis, or unmanned system control. The real physical interval data refers to the upper and lower bounds of observations with clear physical meaning, rather than virtual probability intervals derived from point prediction.

Images