A runoff interval prediction method and system based on a multi-task learning framework

Abstract

The present application belongs to the technical field of runoff prediction, and discloses a runoff interval prediction method and system based on a multi-task learning framework, which comprises the following steps: constructing an interval prediction model based on a multi-task learning framework, taking point prediction values output by a point prediction model of a runoff data sequence and original features as inputs, taking prediction intervals of multiple confidence levels as outputs, and training the interval prediction model according to a total loss function constructed according to quantile loss of the multiple confidence levels; and using the trained interval prediction model to perform runoff interval prediction of the multiple confidence levels. In the interval prediction stage, the multi-task learning mechanism is introduced, the prediction results of multiple quantile points can be output at one time, the high calculation cost and low efficiency problem of the traditional quantile regression which needs to train the model multiple times independently is overcome, the training efficiency is significantly improved, the quantile cross risk is reduced, and a feasible solution is provided for the prediction task with high real-time requirement.

Description

Technical Field

[0001] This invention belongs to the technical field of runoff forecasting, and more specifically, relates to a method and system for predicting runoff intervals based on a multi-task learning framework. Background Technology

[0002] Runoff forecasting is crucial not only for flood control and disaster reduction, and the rational allocation of water resources, but also for meeting the needs of irrigation and power generation. However, the spatiotemporal complexity of runoff sequences presents challenges to accurate runoff forecasting. To improve the accuracy of runoff forecasting, scholars have recently proposed runoff point prediction models based on machine learning methods such as BP neural networks, artificial neural networks, extreme learning machines, and least squares support vector machines (LSSVM), effectively improving the accuracy of runoff forecasting. However, these models mostly focus on point prediction and lack the ability to quantify prediction uncertainty, failing to describe the probability of predicted values ​​and their potential fluctuation ranges. Interval prediction, while providing runoff point prediction results, also depicts the fluctuation range of runoff at a certain confidence level, which is more conducive to the rational allocation of water resources and the formulation of hydropower plans.

[0003] To generate prediction intervals, traditional methods often combine quantile regression with deep learning models. These methods typically train a separate model for each target quantile. For example, to construct prediction intervals at three confidence levels (90%, 80%, and 60%), six independent models need to be trained (each confidence level corresponds to two quantiles above and below). This approach not only significantly increases computational burden and time cost, leading to inefficiency, but also, because the parameters are updated independently between models, it easily causes quantile curve crossing problems—that is, the predicted value of lower quantiles is higher than the predicted value of higher quantiles. This not only affects the theoretical rationality of the prediction intervals but also impairs their reliability and practical utility. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for predicting runoff intervals based on a multi-task learning framework. This method solves the problem that traditional methods for generating prediction intervals based on quantile regression and deep learning models suffer from high computational burden and time costs during model training, resulting in low efficiency. Furthermore, because the parameter updates of each quantile model are independent, the problem of quantile curve intersection is easily caused.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for predicting runoff intervals based on a multi-task learning framework is provided, comprising:

[0006] Training phase:

[0007] An interval prediction model is constructed based on a multi-task learning framework. The point prediction values ​​output by the point prediction model of the runoff data sequence and the original features of the runoff data sequence are used as inputs, and the prediction intervals at multiple confidence levels are used as outputs. The total loss function is constructed based on the quantile loss at multiple confidence levels to train the interval prediction model.

[0008] Prediction phase:

[0009] Point prediction values ​​are generated using a point prediction model based on the original characteristics of the runoff data sequence;

[0010] Based on the point prediction values ​​and the original characteristics of the runoff data sequence, the trained interval prediction model is used to predict runoff intervals at multiple confidence levels.

[0011] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, the interval prediction model includes a shared feature encoding network and a fully connected output layer. The shared feature encoding network is used to receive input data and extract features from the input data. The fully connected output layer is used to receive the extracted features and output prediction intervals with multiple confidence levels. The shared feature encoding network and the fully connected output layer are trained according to the total loss function.

[0012] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, the shared feature encoding network sequentially includes: a sequence input layer for receiving input data; a bidirectional long short-term memory network layer for capturing long-term dependencies from both forward and backward directions and generating feature representations that integrate bidirectional contextual information; a nonlinear activation layer for introducing nonlinear activation to enhance the model's representational ability; and a random deactivation layer for preventing overfitting through a random deactivation mechanism.

[0013] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, the total loss function is a weighted sum of quantile losses at multiple confidence levels; the quantile loss at any confidence level is the sum of the upper quantile loss and the lower quantile loss.

[0014] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, the total loss function Specifically:

[0015] ;

[0016] in, Indicates the first One confidence level; The number of preset confidence levels; For the first Lower quantile loss at each confidence level; For the first Upper quantile loss at each confidence level; These are the actual observed values ​​in the runoff data series; The model predicts the first The lower quantile value of each confidence level; The model predicts the first The upper quantile value at the confidence level.

[0017] The runoff interval prediction method based on a multi-task learning framework provided by the present invention provides a quantile loss at any confidence level. Specifically:

[0018] ;

[0019] When calculating the lower quantile loss, the target quantile level The corresponding quantile levels are as follows:

[0020] ;

[0021] When calculating the upper quantile loss, the target quantile level The corresponding upper quantile levels are as follows:

[0022] ;

[0023] Where, 0 < <1; The lower quantile level; The upper quantile level; Confidence level; These are the actual observed values ​​in the runoff data series; This represents the corresponding quantile value predicted by the model.

[0024] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, the point prediction model sequentially includes: a convolutional layer for extracting local short-term features of the input sequence; a long short-term memory network layer for capturing long-term dependencies to obtain high-level temporal features; and a fully connected layer for integrating high-level temporal features and outputting predicted values.

[0025] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, each data point of the runoff data sequence includes a target variable and multiple feature quantities;

[0026] Accordingly, the point prediction model outputs the following specific point prediction values:

[0027] For each data point, the Pearson correlation coefficient is calculated to select the most relevant key features.

[0028] The normalized runoff data sequence is divided into samples using a sliding window method. Each input sample includes target variable data from the first preset number of historical time steps and key feature data from the second preset number of historical time steps. The output is the predicted value of the target variable for the future time step.

[0029] According to the runoff interval prediction method based on a multi-task learning framework provided by the present invention, the input of the interval prediction model is the point prediction value generated by the point prediction model plus the input sample of the point prediction model.

[0030] According to another aspect of the present invention, a runoff interval prediction system based on a multi-task learning framework is provided, characterized in that the system includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the runoff interval prediction method based on the multi-task learning framework described above.

[0031] In summary, compared with the prior art, the runoff interval prediction method and system based on a multi-task learning framework provided by this invention offer the following advantages:

[0032] 1. In the interval prediction stage, a multi-task learning mechanism is introduced, which can output the prediction results of multiple quantiles at one time. This overcomes the high computational cost and inefficiency of traditional quantile regression, which requires multiple independent training of the model. The training efficiency is significantly improved. Moreover, the output of multiple confidence level prediction intervals shares a set of model parameters, which reduces the risk of quantile crossover. This provides a practical solution for forecasting tasks with high real-time requirements.

[0033] 2. The multi-task learning bidirectional long short-term memory interval prediction model is a deep fusion architecture based on bidirectional long short-term memory network and quantile regression. It can be used to quantify prediction uncertainty and generate prediction intervals with multiple confidence levels. In the interval prediction stage, a multi-task learning framework with hard parameter sharing is adopted. The quantile prediction of multiple confidence intervals is learned synchronously through the shared bidirectional long short-term memory feature extraction layer. All prediction tasks share the same network parameters and are jointly optimized by summing the quantile loss with equal weights.

[0034] 3. In the point prediction stage, a hybrid model of convolutional neural network-long short-term memory network (CNN-LSTM) is used to collaboratively extract local spatial features and long-term temporal dependencies of the input data, thereby improving prediction accuracy;

[0035] 4. When applied to runoff interval prediction tasks, this invention can generate prediction intervals with more reasonable coverage and more compact width, significantly improving the reliability and practicality of uncertainty quantification results. Attached Figure Description

[0036] Figure 1This is a flowchart of the runoff interval prediction method based on a multi-task learning framework provided in an embodiment of the present invention.

[0037] Figure 2 This is a time series comparison of the runoff prediction results of the prediction models at various points provided in the embodiments of the present invention.

[0038] Figure 3 This is a radar chart of the evaluation indicators of the prediction model at each point provided in the embodiments of the present invention.

[0039] Figure 4 This is a radar chart of the evaluation indicators of the prediction models for each interval provided in the embodiments of the present invention at an 80% confidence level.

[0040] Figure 5 This is a comparison chart of the baseline model (CNN-LSTM-BiLSTM) provided in the embodiments of the present invention in the time series of multi-confidence runoff prediction intervals.

[0041] Figure 6 This is a comparison chart of the interval prediction model (CNN-LSTM-MTL-BiLSTM) provided in the embodiments of the present invention in the interval time series of multi-confidence runoff prediction. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0043] Please see Figure 1 This embodiment provides a runoff interval prediction method based on a multi-task learning framework, which includes:

[0044] Training phase:

[0045] An interval prediction model is constructed based on a multi-task learning framework. The point prediction values ​​output by the point prediction model of the runoff data sequence and the original features of the runoff data sequence are used as inputs, and the prediction intervals at multiple confidence levels are used as outputs. The total loss function is constructed based on the quantile loss at multiple confidence levels to train the interval prediction model.

[0046] Prediction phase:

[0047] Point prediction values ​​are generated using a point prediction model based on the original characteristics of the runoff data sequence;

[0048] Based on the point prediction values ​​and the original characteristics of the runoff data sequence, the trained interval prediction model is used to predict runoff intervals at multiple confidence levels.

[0049] This embodiment aims to address the inefficiencies and quantile crossover issues inherent in traditional quantile regression methods. Specifically, it proposes constructing a multi-task learning-based interval prediction model to further quantify prediction uncertainty, building upon runoff data point prediction. This multi-task learning architecture utilizes a hard parameter sharing mechanism to compute multiple target quantiles (corresponding to the upper and lower bounds of prediction intervals at different confidence levels) in parallel at the output layer. A single model then predicts intervals for multiple confidence levels, significantly improving operational efficiency. During the training phase, quantile loss is used as the basic objective. A total loss function is constructed using quantile losses at multiple confidence levels and backpropagated. This allows for joint optimization and implicit regularization at shared feature extraction and fully connected layers. This hard sharing facilitates information transfer between different confidence levels and reduces the risk of quantile crossover.

[0050] In some embodiments, the interval prediction model includes a shared feature encoding network and a fully connected output layer. The shared feature encoding network is used to receive input data and extract features from the input data. The fully connected output layer is used to receive the extracted features and output prediction intervals with multiple confidence levels. The shared feature encoding network and the fully connected output layer are trained according to the total loss function.

[0051] Specifically, the shared feature encoding network includes, in sequence: a sequence input layer for receiving input data; a bidirectional long short-term memory network layer for capturing long-term dependencies from both forward and backward directions to generate feature representations that integrate bidirectional contextual information; a nonlinear activation layer for introducing nonlinear activation to enhance the model's representational ability; and a random deactivation layer for preventing overfitting through a random deactivation mechanism.

[0052] This embodiment proposes a runoff interval prediction method based on a bidirectional long short-term memory (BiLSTM) network using multi-task learning (MTL). By leveraging the BiLSTM structure, it simultaneously captures the forward and backward dependencies of time series data, facilitating a more comprehensive understanding of runoff time-series evolution patterns. Interval prediction requires identifying long-term trends in the data (such as weekly cycles and monthly fluctuations). The gating mechanisms of long short-term memory (forget gate, input gate, output gate) effectively address the "vanishing gradient" problem of ordinary neural networks (RNNs), making it suitable for learning long-term time-series dependencies. Furthermore, BiLSTM further enhances the modeling of "bidirectional long-term associations," providing a richer feature base for multi-task learning.

[0053] The original data and the results of point prediction are used as inputs to the interval prediction model. After the input data enters the model, it first flows through the shared feature encoding network and then passes through the following layers in sequence:

[0054] (1) Sequence Input Layer: serves as the input entry point for the model;

[0055] (2) Bi-Short Memory Network Layer (BiLSTMLayer): As the core temporal feature extractor, it captures long-term dependencies from two directions through forward and backward long short Memory units, respectively, and generates feature representations that integrate bi-directional contextual information;

[0056] (3) Nonlinear activation layer (reluLayer): Introducing nonlinear activation enhances the model's representational ability;

[0057] (4) Dropout Layer: The Dropout mechanism is used to prevent overfitting and improve the generalization ability of the model.

[0058] After being processed by the shared network, the features enter the multi-task output module:

[0059] The fully connected output layer projects the shared temporal representation extracted by BiLSTM onto a high-dimensional task output space, specifically mapped to a dimension of "output dimension (D) × number of confidence levels (K) × 2". Here, "×2" corresponds to the "upper quantile + lower quantile" required for each confidence level. Specifically, the output dimension (D) is the dimension of the point prediction model's output point prediction values. If there are D output dimensions and K confidence levels, the fully connected layer's output dimension is D × 2K (1 × 6 in the specific experiment below, corresponding to 3 confidence levels for univariate output).

[0060] The multi-task output module also includes a custom quantile module (MultiQRLayer): it receives the multidimensional tensor output by the fully connected layer and calculates the preset quantile level (e.g., ...). The quantile loss (equivalent to 0.90 corresponding to the 90th quantile) is calculated. The quantile losses at all confidence levels are aggregated into a single optimization objective by equal weighting, and all network parameters are updated simultaneously through backpropagation.

[0061] Furthermore, the total loss function is a weighted sum of quantile losses at multiple confidence levels; the quantile loss at any confidence level is the sum of the upper quantile loss and the lower quantile loss.

[0062] Specifically, in this embodiment, all quantile losses are aggregated into a total loss function with equal weights. Specifically:

[0063] ;

[0064] in, Indicates the first One confidence level; The number of preset confidence levels; For the first Lower quantile loss at each confidence level; For the first Upper quantile loss at each confidence level; These are the actual observed values ​​in the runoff data series; The model predicts the first The lower quantile value of each confidence level; The model predicts the first The upper quantile value at the confidence level.

[0065] Among them, the quantile loss at any confidence level Specifically:

[0066] ;

[0067] When calculating the lower quantile loss, the target quantile level The corresponding quantile levels are as follows:

[0068] ;

[0069] When calculating the upper quantile loss, the target quantile level The corresponding upper quantile levels are as follows:

[0070] ;

[0071] Where, 0 < <1; The lower quantile level; The upper quantile level; Confidence level; These are the actual observed values ​​in the runoff data series; This represents the corresponding quantile value predicted by the model.

[0072] In some embodiments, the point prediction model sequentially includes: a convolutional layer for extracting local short-term features of the input sequence; a long short-term memory network layer for capturing long-term dependencies to obtain high-level temporal features; and a fully connected layer for integrating high-level temporal features and outputting predicted values.

[0073] In this embodiment, considering that convolutional neural networks can effectively capture local features and short-term dependencies in temporal data, while long short-term memory networks excel at modeling long-term temporal dependencies, a point prediction model based on a convolutional neural network-long short-term memory network (CNN-LSTM) is proposed. By effectively extracting local features from sequence data through CNN layers and then combining them with LSTM layers to capture long-term temporal dependencies, the accuracy of point prediction can be significantly improved, generating high-precision deterministic predictions.

[0074] Specifically, the preprocessed runoff sequence data is first input, and local short-term features are extracted from the input sequence through a one-dimensional convolutional layer. To accelerate network training and improve stability, a batch normalization layer is added immediately after the convolutional layer to standardize the activation values ​​of the convolutional output. Then, a ReLU activation function and a max-pooling layer with a stride of 1 are used for nonlinear transformation and feature dimensionality reduction. The resulting feature sequence is then input into an LSTM layer to capture long-term dependencies. Finally, a fully connected layer is used to integrate high-level temporal features and output predicted values. The model uses the Adam optimizer with mean squared error as the loss function, sets the batch size and maximum number of training epochs, and employs a piecewise decay strategy for learning rate during training.

[0075] Furthermore, before the point prediction model performs runoff point value prediction, it also includes loading and preprocessing the runoff sequence data. Loading the runoff time series dataset, i.e., the runoff data sequence, wherein each data point in the runoff data sequence includes a target variable and multiple feature quantities; the multiple feature quantities include the dew point temperature at a height of 2 meters, d 2m 2 meters temperature, t 2m Evaporation, e, U component of wind at a height of 10 meters, u 10 The V component of wind at a height of 10 meters, v 10 And surface runoff, multiple in sro; the target variable is the inflow, Q.

[0076] Accordingly, the training of the point prediction model, which outputs the predicted point value, specifically includes:

[0077] 80% of the dataset was used as the training set, 10% as the validation set, and 10% as the test set. For each data point, Pearson correlation coefficients were calculated to identify the most relevant key features; these key features were the meteorological factors most relevant to the inflow.

[0078] To improve the model's generalization ability and stability, the sequence data is normalized, with the normalized data falling within the interval (0,1). The normalized runoff data sequence is then divided into samples using a sliding window method. Each input sample includes target variable data from a first preset number of historical time steps (e.g., 7 time steps) and key feature data from a second preset number of historical time steps (e.g., 4 selected meteorological factors). The output is the predicted value of the target variable (i.e., inflow) for the next time step (e.g., 8th time step). The sliding window size is set to 1 to generate continuous sample pairs.

[0079] Furthermore, the input to the interval prediction model is the point prediction value generated by the point prediction model plus the input sample of the point prediction model. The output of the interval prediction model needs to go through a post-processing layer: namely, inverse normalization and interval generation. The model predicts "standardized quantiles", which are then restored to the true scale using Z-score standardization.

[0080] Furthermore, this embodiment also includes model prediction and evaluation during the training phase. The performance differences between related methods are evaluated, and finally, a comprehensive evaluation metric is used to systematically verify the accuracy of point prediction and the reliability of interval prediction, providing a comprehensive and reliable solution for prediction.

[0081] To comprehensively evaluate the predictive performance and computational efficiency of the experimental design model, this embodiment uses five point prediction and five interval prediction evaluation metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (CDE). The following metrics are used: Coverage Probability (PICP); Mean Prediction Interval Width (PINAW); Coverage Width-based Criterion (CWC); Mean Prediction Interval Center Deviation (MPICD); and Average Interval Score (AIS).

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] In the formula, It is the first i The true value of each sample; It is the first i Predicted values ​​for each sample; 、 Predict the lower and upper limits of the interval respectively; y max The maximum value of the true value; y min It is the minimum value of the true value; The significance level; γ , η This is a hyperparameter that controls the "penalty strength when coverage deviates from the expected value"; S (a) ( y i ) is the interval score function; n For the length of the test set, I (·) is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0093] In other embodiments, a runoff interval prediction system based on a multi-task learning framework is also provided. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the runoff interval prediction method based on the multi-task learning framework described above.

[0094] The purpose of this embodiment is to propose a runoff interval prediction method based on a multi-task learning bidirectional long short-term memory (BiLSTM) network. This method first uses convolutional layers to extract local features from the input sequence, and then uses a long short-term memory network layer to model long-term dependencies, outputting accurate point prediction values. Subsequently, these prediction values ​​and the original features are input into the multi-task BiLSTM interval prediction model. The bidirectional structure of the BiLSTM network captures temporal forward and backward dependencies, and multiple quantile outputs are optimized in parallel through a multi-task learning mechanism to generate prediction intervals with different confidence levels in a single step. Experimental results verify that the proposed method significantly improves both point prediction accuracy and interval prediction reliability, providing an effective solution for time-series prediction tasks requiring accurate assessment of uncertainty.

[0095] In some specific examples, the hydrological data of Xiaowan Hydropower Station in Yunnan Province are used as an example to verify the effectiveness of the present invention.

[0096] The example uses a daily timescale to obtain daily inflow (Q) data for the Xiaowan Hydropower Station from January 1, 2013 to December 31, 2023, as well as dew point temperature (d) at a height of 2 meters. 2m ), 2-meter temperature (t) 2m ), evaporation (e), U component of 10-meter wind speed (u) 10 ), the v component of the 10-meter wind speed (v 10 Meteorological data on surface runoff (SRO) were collected, and missing values ​​were imputed. All data were then normalized to fall within the (0,1) interval to eliminate the influence of different units of measurement. The complete dataset was proportionally divided into a training set (80%), a validation set (10%), and a test set (10%).

[0097] A sliding window method was used to convert time series data into a sample format suitable for training the prediction model. This included historical inflow data from the previous 7 days and meteorological characteristics from the previous 2 days (based on four key meteorological factors after screening). 2m t 2m , e, u 10 Using the inflow rate on day 8 as the prediction target, a sliding step size of 1 is set to construct a continuous sample sequence for subsequent modeling training.

[0098] A point prediction model (CNN-LSTM) is constructed for point prediction. First, preprocessed sequence data is input, and local short-term features are extracted from the input sequence through a one-dimensional convolutional layer (equipped with 16 kernels of size 2). To accelerate network training and improve stability, a batch normalization layer is added immediately after the convolutional layer to standardize the activation values ​​of the convolutional output. Then, a ReLU activation function and a max-pooling layer with a stride of 1 are used for non-linear transformation and feature dimensionality reduction. The resulting feature sequence is then input into an LSTM layer with two hidden layers (each containing 16 neurons) to capture long-term dependencies. Finally, a fully connected layer integrates high-level temporal features and outputs the predicted point values. The model uses the Adam optimizer with mean squared error as the loss function, a batch size of 50, a maximum training epoch of 100, and a piecewise decay strategy for learning rate during training.

[0099] The obtained point prediction results are fused with historical meteorological feature data and used as input for the interval prediction model in this stage. The core of the model uses a bidirectional LSTM layer to simultaneously capture the forward and backward long-term dependencies of the fused feature sequence; and employs a multi-task learning framework, sharing the underlying network parameters to compute multiple target quantiles in parallel at the output layer, corresponding to the upper and lower bounds of intervals at different confidence levels. During the training phase, the quantile loss function is used as the optimization objective, directly optimizing different quantiles to simultaneously ensure the reliability and accuracy of the prediction intervals. The model uses the Adam optimizer, combined with a gradient pruning strategy to ensure training stability.

[0100] The predicted point values ​​and upper and lower bounds of the prediction intervals output by the model are denormalized to restore them to the scale of the original traffic data. To comprehensively evaluate the effectiveness of the proposed prediction model, the experiments are conducted from two dimensions: point prediction accuracy and interval prediction reliability. For point prediction, Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Decision Tree Models (DT), and the proposed method (CNN-LSTM) are compared. For interval prediction, the CNN-LSTM-BiLSTM interval prediction model is used as the baseline model, and the CNN-LSTM-MTL-BiLSTM interval prediction model is used as the method proposed in this invention for comparison. The baseline model does not introduce a multi-task learning mechanism and requires multiple BiLSTM network models. A separate model is trained for each target quantile (corresponding to the upper and lower bounds of the interval at different confidence levels). The model structures are the same, but the parameters are completely independent. The evaluation system covers 5 point prediction metrics and 5 interval prediction metrics to comprehensively evaluate the model performance from multiple perspectives. For example, the prediction results are... Figure 2 , Figure 3 , Figure 4 As shown, the quantitative calculation results are shown in Tables 1 and 2.

[0101] Table 1 Comparison of performance evaluation metrics for prediction models at different points

[0102]

[0103] Table 2. Comparison of performance evaluation indicators of prediction models for each interval at the 80% confidence level.

[0104]

[0105] Comprehensive analysis shows that the point prediction hybrid model provided by this invention performs best among all point prediction comparison models, with the lowest error indices (MAE, MSE, RMSE, MAPE) and the best goodness of fit (…). The highest value was achieved, verifying its ability to effectively capture nonlinear characteristics in hydrological time series. (Reference) Figure 4 , Figure 5 and Figure 6In terms of interval prediction, taking the 80% confidence interval as an example, the interval prediction model provided by this invention outperforms the benchmark model on all interval evaluation indicators, generating more reliable, compact, and accurate prediction intervals. This demonstrates that the multi-task learning framework significantly improves the ability to quantify prediction uncertainty by jointly optimizing multiple quantiles. Overall, the interval prediction method based on the CNN-LSTM-MTL-BiLSTM model exhibits superior performance in runoff forecasting for the Xiaowan Hydropower Station in Yunnan Province, providing reliable technical support for the refined scheduling and risk decision-making of hydropower stations.

[0106] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and 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 runoff interval prediction method based on a multi-task learning framework, characterized in that, include: Training phase: An interval prediction model is constructed based on a multi-task learning framework. The point prediction values ​​output by the point prediction model of the runoff data sequence and the original features of the runoff data sequence are used as inputs, and the prediction intervals at multiple confidence levels are used as outputs. The total loss function is constructed based on the quantile loss at multiple confidence levels to train the interval prediction model. Prediction phase: Point prediction values ​​are generated using a point prediction model based on the original characteristics of the runoff data sequence; Based on the point prediction values ​​and the original characteristics of the runoff data sequence, the trained interval prediction model is used to predict runoff intervals at multiple confidence levels. The interval prediction model includes a shared feature encoding network and a fully connected output layer. The shared feature encoding network is used to receive input data and extract features from the input data. The fully connected output layer is used to receive the extracted features and output prediction intervals with multiple confidence levels. The shared feature encoding network and the fully connected output layer are trained according to the total loss function. The total loss function is a weighted sum of quantile losses at multiple confidence levels; the quantile loss at any confidence level is the sum of the upper quantile loss and the lower quantile loss.

2. The runoff interval prediction method based on a multi-task learning framework as described in claim 1, characterized in that, The shared feature encoding network comprises, in sequence: a sequence input layer for receiving input data; a bidirectional long short-term memory network layer for capturing long-term dependencies from both forward and backward directions and generating feature representations that integrate bidirectional contextual information; a nonlinear activation layer for introducing nonlinear activation to enhance the model's representational ability; and a random deactivation layer for preventing overfitting through a random deactivation mechanism.

3. The runoff interval prediction method based on a multi-task learning framework as described in claim 1, characterized in that, The total loss function Specifically: ； in, Indicates the first One confidence level; The number of preset confidence levels; For the first Lower quantile loss at each confidence level; For the first Upper quantile loss at each confidence level; These are the actual observed values ​​in the runoff data series; The model predicts the first The lower quantile value of each confidence level; The model predicts the first The upper quantile value at the confidence level.

4. The runoff interval prediction method based on a multi-task learning framework as described in claim 1, characterized in that, Quantile loss at any confidence level Specifically: ； When calculating the lower quantile loss, the target quantile level The corresponding quantile levels are as follows: ； When calculating the upper quantile loss, the target quantile level The corresponding upper quantile levels are as follows: ； Where, 0 < <1; The lower quantile level; The upper quantile level; Confidence level; These are the actual observed values ​​in the runoff data series; This represents the corresponding quantile value predicted by the model.

5. The runoff interval prediction method based on a multi-task learning framework as described in any one of claims 1-4, characterized in that, The point prediction model comprises, in sequence: a convolutional layer for extracting local short-term features of the input sequence; a long short-term memory network layer for capturing long-term dependencies to obtain high-level temporal features; and a fully connected layer for integrating high-level temporal features and outputting predicted values.

6. The runoff interval prediction method based on a multi-task learning framework as described in any one of claims 1-4, characterized in that, Each data point in the runoff data sequence includes a target variable and multiple feature quantities; Accordingly, the point prediction model outputs the following specific point prediction values: For each data point, the Pearson correlation coefficient is calculated to select the most relevant key features. The normalized runoff data sequence is divided into samples using a sliding window method. Each input sample includes target variable data from the first preset number of historical time steps and key feature data from the second preset number of historical time steps. The output is the predicted value of the target variable for the future time step.

7. The runoff interval prediction method based on a multi-task learning framework as described in claim 6, characterized in that, The input to the interval prediction model is the point prediction value generated by the point prediction model plus the input sample of the point prediction model.

8. A runoff interval prediction system based on a multi-task learning framework, characterized in that, The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the runoff interval prediction method based on the multi-task learning framework as described in any one of claims 1-7.

Images