A method for predicting land surface temperature and evaporation

Abstract

This invention relates to the field of meteorological forecasting technology and discloses a method for predicting surface temperature and evaporation. The invention first performs multi-scale mode decomposition on the raw meteorological data, separating high-frequency noise components from mid- and low-frequency effective information components. This eliminates interference from high-frequency random disturbances at the data preprocessing level, effectively solving the inherent non-stationarity and strong volatility problems of meteorological data. Subsequently, a CNN-BiLSTM-Attention hybrid model is constructed to perform parallel predictions on the selected mode components. The core of this invention lies in introducing a barrel theory optimization algorithm to adaptively and jointly optimize the mode decomposition parameters and deep learning hyperparameters, overcoming the limitations of manual parameter tuning. Finally, the prediction results of each component are reconstructed, and long-term predicted values ​​are output. This effectively solves the problems of insufficient accuracy, difficulty in parameter tuning, and weak cross-site generalization ability of traditional methods and existing combined models when dealing with non-stationary, multi-scale meteorological sequences, significantly improving the accuracy, stability, and automation level of prediction.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method for predicting surface temperature and evaporation. Background Technology

[0002] Surface temperature and evaporation are core physical quantities characterizing regional hydrothermal cycles and surface energy balance, and are crucial for drought monitoring, agricultural irrigation, water resource management, and climate change research. Traditional forecasting methods are mainly divided into two categories: the first is models based on physical mechanisms, such as the Penman-Monteith model. While these models have clear physical meaning, they heavily rely on various meteorological and underlying surface data, such as wind speed, radiation, and vegetation parameters. Parameter acquisition is costly, regional adaptability is poor, and accuracy drops sharply when faced with missing data or changes in the climate background. The second category is models based on statistics and classical time series, such as ARIMA. These models typically rely on the assumption of time series stationarity, making it difficult to effectively characterize the prevalent nonlinearity, non-stationarity, and multi-period superposition characteristics in meteorological data. In long-term forecasts, error accumulation is severe.

[0003] In recent years, deep learning methods have been gradually applied to meteorological element prediction. The typical workflow of traditional prediction methods usually includes: first, collecting meteorological observation sequences from multiple stations and performing quality control on missing and outlier values; second, constructing input features and forming training samples using a sliding window; then, directly inputting the samples into networks such as LSTM, GRU, or CNN for end-to-end training, and finally outputting multi-step prediction results. While this workflow can improve the fitting ability to nonlinear relationships, its key drawback lies in the direct modeling of the original sequences. Meteorological sequences generally contain high-frequency noise, multi-scale fluctuations, and abrupt changes. Network training is easily influenced by redundant perturbations, leading to amplified fluctuations and insufficient stability in the prediction results. Furthermore, the single... Networks struggle to distinguish and utilize the information carried by different frequency components. When migrating across sites, models are more sensitive to distribution differences, limiting their generalization ability and interpretability. These problems are more pronounced in arid and semi-arid regions because precipitation is scarce but highly event-driven, surface energy balance is significantly driven by radiation and wind fields, and evaporation processes exhibit sudden and strongly nonlinear characteristics. This makes surface temperature and evaporation sequences more prone to spikes, abrupt changes, and multi-scale coupled fluctuations. Traditional end-to-end training processes are less likely to obtain stable and reliable long-term prediction results. Therefore, it is necessary to carry out targeted improvements for the complex sequence characteristics of arid and semi-arid regions and to select relevant sites as preferred application scenarios to reflect the engineering requirements of robust long-term prediction methods.

[0004] To improve the sensitivity of end-to-end modeling to nonstationarity, existing technologies have further developed a combined prediction process of "decomposition, prediction, and reconstruction." First, EMD or VMD is used to perform modal decomposition on the original sequence to obtain several intrinsic modal components and residual terms. Then, deep learning models are trained on each component for parallel prediction. Finally, the prediction results are reconstructed to obtain the final output. This process reduces the complexity of the sequence through decomposition, but it also exposes new bottlenecks: the decomposition quality and prediction performance are highly dependent on the decomposition parameters (such as the number of modes, penalty factor, etc.) and network hyperparameters (such as the number of layers, learning rate, etc.). In engineering practice, parameter combinations are often determined by manual experience, trial and error, or grid search. The parameter tuning cost is high and the reproducibility is poor. Moreover, it is difficult to maintain a uniform optimal configuration for different sites and different prediction durations, which easily leads to local optima and insufficient generalization ability. Especially in arid and semi-arid regions, when the sequence is simultaneously affected by extreme weather processes and sudden changes in the underlying surface dryness and wetness, small parameter deviations can cause unstable mode division and amplified error accumulation, further weakening the reliability of cross-site migration and long-term extrapolation. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting land surface temperature and evaporation.

[0006] This invention provides a method for predicting land surface temperature and evaporation, comprising the following steps: Meteorological data from meteorological stations are collected, including surface temperature, evaporation, and meteorological factor sequences. A hierarchical preprocessing strategy is used to perform missing value imputation and outlier removal preprocessing on the meteorological data. A hybrid prediction model is constructed, which integrates convolutional neural networks, bidirectional long short-term memory networks, and attention mechanisms to model the nonlinear, high-dimensional, and time-series correlation features of surface temperature, evaporation, and meteorological factor sequences. The original surface temperature sequence and evaporation sequence were decomposed into multiple intrinsic mode function components of different frequencies using the mode decomposition method. The barrel theory optimization algorithm is introduced to adaptively and jointly optimize the mode decomposition parameters and the hyperparameters of the hybrid prediction model, so as to obtain the optimized fusion prediction model. The fusion prediction model is used to make multi-step predictions of surface temperature and evaporation to obtain predicted values ​​within a preset time range.

[0007] Furthermore, the construction of the hybrid prediction model includes: Local features are extracted from the original meteorological factor sequence using a convolutional neural network to identify local fluctuation patterns, short-term trends, and correlations in the input data. By using a bidirectional long short-term memory network to perform time-series modeling on the feature sequences extracted by the convolutional neural network, the long-term dependence of meteorological factors over time and bidirectional dynamic information can be captured. By calculating the weight distribution at different time steps through the attention layer, the system focuses on the part of the historical sequence that has the greatest influence on the current prediction, and generates a weighted global feature representation.

[0008] Furthermore, the mode decomposition method includes two approaches: empirical mode decomposition and variational mode decomposition. Both empirical mode decomposition and variational mode decomposition perform multi-scale decomposition on the surface temperature sequence and evaporation sequence, extracting intrinsic mode function components of different frequencies. By comparing and analyzing the decomposition stability, anti-mode aliasing ability, and subsequent prediction accuracy of the two methods, the method with higher decomposition stability and better prediction performance is selected for subsequent prediction.

[0009] Furthermore, the barrel theory optimization algorithm treats the mode decomposition parameters and the hyperparameters of the hybrid prediction model as unified variables to be optimized, and achieves a dynamic balance between global search and local convergence through barrel adjustment and perturbation search mechanisms.

[0010] Furthermore, the multi-step forecast includes predictions of surface temperature and evaporation over the next 30, 90, and 180 days.

[0011] Furthermore, it also includes statistical analysis of the prediction loss of each intrinsic mode function component after mode decomposition, and retaining the first few order mode components with prediction errors below a preset threshold for hybrid prediction model modeling.

[0012] Furthermore, it also includes weight distribution analysis based on attention mechanism to determine the influence weights of different meteorological factors on the prediction results of surface temperature and evaporation, revealing the nonlinear regulatory relationship between meteorological factors.

[0013] Furthermore, in the surface temperature prediction, the influence weights of each meteorological factor from high to low are as follows: air temperature, sunshine duration, evaporation, relative humidity, air pressure, and rainfall; In the evaporation forecast, the influence weights of each meteorological factor, from highest to lowest, are: air temperature, surface temperature, sunshine duration, relative humidity, air pressure, and precipitation.

[0014] Furthermore, the method supports simultaneous prediction of time-series data from multiple monitoring stations and multiple meteorological elements.

[0015] A surface temperature and evaporation prediction system, comprising: The data collection and preprocessing module is used to collect meteorological data from multiple stations and perform missing value imputation and outlier removal. The hybrid prediction model module includes a convolutional neural network layer, a bidirectional long short-term memory network layer, and an attention mechanism layer, which are used to model and predict each modal component separately. The mode decomposition module is used to decompose the original surface temperature sequence and evaporation sequence into multiple intrinsic mode function components; The optimization module is used to introduce the barrel theory optimization algorithm to adaptively and jointly optimize the mode decomposition parameters and the hyperparameters of the hybrid prediction model. The prediction execution module is used to perform multi-step predictions of surface temperature and evaporation using the optimized fusion prediction model. The weighting analysis module is used to analyze the influence weights of different meteorological factors on the prediction results based on the attention mechanism.

[0016] Compared with the prior art, the technical solution provided by the embodiments of the present invention has the following advantages: The present invention first performs multi-scale mode decomposition on the original meteorological data, separating the high-frequency noise components and the mid-to-low-frequency effective information components in the data, eliminating the interference of high-frequency random disturbances from the data preprocessing level, and effectively solving the problem of non-stationarity and strong fluctuation inherent in meteorological data. Then, a CNN-BiLSTM-Attention hybrid model is constructed to perform parallel prediction of each selected mode component. Its core lies in introducing the barrel theory optimization algorithm to adaptively and jointly optimize the mode decomposition parameters and deep learning hyperparameters to overcome the limitations of manual parameter tuning. Finally, the prediction results of each component are reconstructed and long-term prediction values ​​are output. This effectively solves the problems of insufficient accuracy, difficulty in parameter tuning, and weak cross-site generalization ability of traditional methods and existing combined models when dealing with non-stationary, multi-scale meteorological sequences, and significantly improves the accuracy, stability and automation level of prediction. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method for predicting land surface temperature and evaporation. Figure 2 This is a schematic diagram of the CNN-BiLSTM-Attention fusion structure; Figure 3 This is a schematic diagram of the mode decomposition framework; Figure 4 Optimization strategy diagram for the barrel; Figure 5 This is a plot of the prediction loss for the modal components; Figure 6 This is a diagram showing the long-term prediction results of surface temperature and evaporation from the BTO-EMD combined model. Figure 7 This is a diagram showing the long-term prediction results of surface temperature and evaporation from the BTO-VMD combined model. Detailed Implementation

[0018] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0021] Example 1 like Figures 1-7 As shown, this embodiment discloses a method for predicting land surface temperature and evaporation, aiming to solve the technical problems in existing long-term prediction of land surface temperature and evaporation, such as the difficulty of adapting the models to the nonlinearity, non-stationarity, and cross-site differences of multi-source meteorological sequences, resulting in low prediction accuracy, poor robustness, and insufficient model stability due to the reliance on manual experience in setting mode decomposition parameters and network hyperparameters. The specific implementation steps are as follows: S1: Meteorological Data Collection and Preprocessing This embodiment selects three typical meteorological stations (Pingluo Station, Shizuishan Station, and Yinchuan Station) in the arid and semi-arid regions of Northwest my country as research objects, covering different topographical features (plains, valleys, and suburbs) to improve the model's cross-site generalization ability and adapt to the forecasting needs of different regions.

[0022] Data collection: Daily meteorological observation data from the three stations mentioned above were collected through the China Meteorological Data Network and local meteorological observation stations, totaling 3,652 sets of valid data (excluding redundant data from leap years). Data types include: surface temperature (GST, unit: °C), evaporation (Evap, unit: mm), air temperature (unit: °C), sunshine duration (unit: h), relative humidity (unit: %), air pressure (unit: hPa), and precipitation (unit: mm). Among them, surface temperature and evaporation are used as target variables for prediction, and the other meteorological factors are used as input feature variables.

[0023] Data preprocessing: To address missing values, outliers, and noise in the collected raw meteorological data, a hierarchical preprocessing strategy was adopted to ensure data quality and provide a reliable foundation for subsequent model training. (1) Missing value imputation: A hybrid method of linear interpolation and neighbor mean imputation is used to imput a small amount of missing data (missing rate <5%) for a single day at a single station; for missing data for 3 consecutive days or more (missing rate 5%-10%), mean imputation is performed by combining the historical mean of the same station and the data of neighboring stations for the same period, to ensure that the imputed data conforms to the actual meteorological change pattern and avoids the deviation caused by a single interpolation method; (2) Outlier removal: The 3σ criterion is used to identify outliers. The mean μ and standard deviation σ of each meteorological variable are calculated. Data that exceed the range of [μ-3σ, μ+3σ] are identified as outliers. Combined with the meteorological trend of the same station during the same period, linear interpolation of nearby normal data is used to replace outliers to avoid interference of outliers with model training. (3) Data standardization: The Min-Max standardization method is used to normalize all input feature variables and target variables to the [0,1] interval to eliminate the influence of differences in the units of different variables. The standardization formula is as follows: , In the formula, The data is standardized, and x represents the original data. The minimum value of this variable. This is the maximum value of the variable; (4) Data set partitioning: The preprocessed complete dataset is divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for training model parameters, the validation set is used to adjust model hyperparameters and avoid overfitting, and the test set is used to evaluate the final prediction performance of the model. At the same time, in order to verify the long-term prediction ability of the model, multi-step prediction datasets for the next 30 days, 90 days and 180 days are constructed respectively. The input sequence is constructed using the sliding window method, and the window length is set to 60 days (that is, using the meteorological data of the previous 60 days to predict the surface temperature and evaporation of the next 30 days, 90 days and 180 days).

[0024] S2: Constructing a CNN-BiLSTM-Attention hybrid prediction model To address the challenges of traditional deep learning models in capturing the cross-scale fluctuations and cross-site differences in multi-source meteorological sequences, as well as the unreasonable weight allocation of key features, a structured enhancement improvement is implemented on the CNN-BiLSTM-Attention model to construct an improved hybrid model for long-term prediction of cross-site surface temperature and evaporation. The specific improvements are as follows: (1) Improvement of multi-scale temporal convolution feature extraction end: A multi-scale temporal convolution structure is introduced, and three sets of convolution kernels of different sizes (3×1, 5×1, 7×1) are set to perform convolution operations on the input original meteorological factor sequence in parallel, capturing the temporal scale fluctuation features of short-term (within 3 days), medium-term (within 5 days), and long-term (within 7 days) respectively; then, through feature splicing and batch normalization processing, the multi-scale convolution features are fused to suppress the interference of high-frequency noise on key trend information and improve the completeness and reliability of feature extraction. The output formula of the CNN feature extraction layer is as follows: , In the formula: The output of the j-th convolutional kernel in layer l at time t; This is the input of the i-th feature from the previous layer at time t; Let be the convolution kernel corresponding to the ij-th layer in the l-th layer; For bias terms; It is the ReLU activation function; (2) Improvement of time series modeling end: an adaptive feature calibration module for cross-site distribution differences is introduced. This module calculates the mean and variance of the input features of each site, constructs adaptive calibration coefficients, and dynamically calibrates the feature vectors of different sites. This enables the network to maintain stable feature representation ability when processing data from different sites (where there are differences in feature distribution due to terrain and climate differences), and avoids the decrease in accuracy when predicting across sites. (3) BiLSTM time series modeling: The BiLSTM layer consists of a forward LSTM and a backward LSTM. The forward LSTM captures the forward evolution characteristics of the time series, and the backward LSTM captures the backward dependency characteristics of the time series. By splicing the outputs of the two, the temporal correlation of the meteorological series is fully characterized, which is suitable for the long-term temporal evolution of surface temperature and evaporation. The specific calculation formula of the BiLSTM layer is as follows: , , In the formula: The CNN feature vectors within time step t; , Forward and inverse LSTM outputs; The concatenated BiLSTM output is used for subsequent attention processing; (4) Improved Attention Mechanism: The traditional attention mechanism, which only weights the time dimension, is extended to a joint weighting of the time dimension and the variable / modal dimension. It not only focuses on the key meteorological information at different time steps, but also emphasizes the importance of different input features (meteorological factors) and different modal components, so as to achieve accurate fusion of key meteorological factors and effective modal information, and improve the robustness, generalization ability and interpretability of the model in long-term prediction tasks. The specific calculation formula of the Attention layer is as follows: , , , In the formula: Let be the hidden state of BiLSTM at time step t; To train the weight matrix; These are training vectors; For bias terms; is the attention weight at time step t; c is the context vector, representing the weighted global feature representation, which serves as the input to the final output prediction layer; (5) Initial model parameter settings: 3 convolutional layers, with 64, 128, and 256 kernels per group, kernel sizes of 3×1, 5×1, and 7×1 respectively, stride of 1, and padding mode of "same"; 2 BiLSTM layers, with an initial number of hidden units of 128 per layer, and dropout coefficient of 0.3 (to prevent overfitting); attention layer weight matrix... Initialized as a random normal matrix, bias term Initialize to 0; the output layer uses a linear activation function with an output dimension of 2 (corresponding to the predicted values ​​of surface temperature and evaporation, respectively); the optimizer uses the Adam optimizer, with an initial learning rate of 0.001, an initial number of iterations of 100, and a batch size of 32.

[0025] S3: Model Fusion, Long-Term Prediction, and Explanation of Mechanisms of Action To enhance the model's adaptability to features such as non-stationarity, multi-frequency disturbances, and trend noise in meteorological time series, a mode decomposition strategy and the barrel theory optimization algorithm (BTO) are introduced based on the CNN-BiLSTM-Attention model constructed in step S2 to build a fusion model. This enables long-term accurate prediction of surface temperature and evaporation, and explains the mechanism of action of meteorological factors. The specific steps are as follows: S3.1: Original Sequence Mode Decomposition Two typical mode decomposition methods, Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD), were used to perform multi-scale decomposition on the preprocessed raw sequences of land surface temperature and evaporation from three stations, extracting intrinsic mode functions (IMFs) and residual terms (res) at different frequencies and feature levels. The initial settings of the specific decomposition parameters are as follows: (1) EMD decomposition parameters: The termination condition is set to the standard deviation of IMF being less than 0.2, the maximum number of decomposition modes is set to 8, the endpoint effect is suppressed by mirror extension method, and the original nonlinear and nonstationary meteorological sequence is adaptively decomposed into several IMF components with local characteristics (high frequency components correspond to short-term meteorological fluctuations, and low frequency components correspond to long-term climate trends) and 1 residual term (corresponding to the overall trend of the sequence). (2) VMD decomposition parameters: the penalty factor α is initially set to 2000, the noise tolerance τ is set to 1e-7, the maximum number of iterations is set to 1000, and the number of decomposition modes K is initially set to 8. The signal is stably decomposed and denoised through the variational optimization framework to avoid the mode aliasing problem that is easy to occur in EMD decomposition and improve the stability and consistency of the decomposed components.

[0026] After decomposition, the IMF components obtained by the two methods are initially screened to remove invalid components: by statistically analyzing the prediction loss distribution of each IMF component, it is found that the prediction error of the fourth and subsequent IMF components (imf4 and subsequent ones) of both EMD and VMD has dropped to within 0.02, which is difficult to provide effective prediction information and will increase the computational cost of the model. Therefore, only the first four IMF components (imf0-imf3) and one residual term are retained for subsequent modeling to balance prediction accuracy and computational efficiency.

[0027] S3.2: Adaptive Parameter Optimization of BTO Algorithm To address the issues of fluctuating decomposition quality, unstable model training, and poor cross-site generalization ability caused by the reliance on manual experience in setting modality decomposition parameters and model hyperparameters, a Barrel Theory Optimization (BTO) algorithm is introduced as an additional optimization module to achieve adaptive optimization of key parameters of the fusion model. The specific optimization process is as follows: (1) Constructing the joint parameter vector: The key decomposition parameters of VMD / EMD and the key hyperparameters of the improved CNN-BiLSTM-Attention model are combined to form a joint parameter vector X, where: VMD decomposition parameters: penalty factor α (range: 1000-5000), number of decomposition modes K (range: 4-8); EMD decomposition parameters: maximum number of decomposition modes (value range: 4-8), endpoint extension length (value range: 5-15). CNN-BiLSTM-Attention hyperparameters: number of convolutional kernels (range: 64-256), number of BiLSTM hidden units (range: 64-256), attention dimension (range: 32-128), learning rate (range: 1e-4-1e-2), dropout coefficient (range: 0.1-0.5). (2) Set the BTO algorithm parameters: set the population size to 30, the maximum number of iterations to 50, the fitness function to the average R² value of the model's prediction of surface temperature and evaporation on the validation set (the larger the R² value, the better the model performance), and the optimization objective is to maximize the fitness function value; (3) Adaptive optimization process: ① Initialize the population and randomly generate 30 sets of joint parameter vectors; ② For each set of parameter vectors, the original sequence is decomposed using the corresponding mode decomposition method (EMD / VMD) to obtain the filtered IMF components and residual terms; ③ Input each component into the improved CNN-BiLSTM-Attention model for training, and obtain the fitness value of the model on the validation set; ④ Based on the fitness value, the joint parameter vector is updated through the selection, crossover, and mutation operations of the BTO algorithm, and iterative optimization is performed; ⑤ When the number of iterations reaches the maximum value, or the fitness value tends to stabilize (the fitness value changes by less than 1e-4 in 5 consecutive iterations), stop the optimization and output the optimal parameter combination.

[0028] Through the above optimization process, the optimal fusion models for EMD and VMD are obtained respectively: BTO-EMD-CNN-BiLSTM-Attention model and BTO-VMD-CNN-BiLSTM-Attention model. The optimal parameter combination of the BTO-VMD fusion model is as follows: VMD penalty factor α=2500, decomposition modality number K=6 (imf0-imf3 are retained after filtering); CNN convolutional kernel number is 128, 256, 256, BiLSTM hidden unit number is 128, attention dimension is 64, learning rate is 0.0015, dropout coefficient is 0.3.

[0029] S3.3: Long-term forecasting and outcome reconstruction The optimal parameter combination obtained in step S3.2 is applied to the fusion model to model and predict the test set data of the three sites respectively. The specific process is as follows: (1) The original sequence of surface temperature and evaporation in the test set was decomposed into four IMF components (imf0-imf3) and one residual term using the optimal parameter mode decomposition method (EMD / VMD). (2) Input each IMF component and residual term into the improved CNN-BiLSTM-Attention model to perform separate modeling and multi-step prediction (predict the component values ​​for the next 30 days, 90 days and 180 days respectively). (3) The linear superposition method is used to reconstruct the prediction results of each component into the final predicted values ​​of surface temperature and evaporation. The reconstruction formula is as follows: , In the formula, This is the final predicted value (surface temperature or evaporation). Let be the predicted value of the i-th IMF component. This represents the predicted value of the residual term.

[0030] After the forecasts were completed, the normalized forecast values ​​were converted into actual physical quantities through inverse standardization, yielding the predicted surface temperature and evaporation for the next 30, 90, and 180 days at the three stations. Comparisons between some of the predicted results and the measured values ​​are shown below. Figure 5 , Figure 6 As shown.

[0031] S3.4: Model Performance Evaluation The predictive performance of the two fusion models, BTO-EMD-CNN-BiLSTM-Attention and BTO-VMD-CNN-BiLSTM-Attention, was evaluated using four commonly used evaluation metrics: coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The calculation formulas for the evaluation metrics are as follows: , , , , In the formula, These are measured values. For predicted values, The mean of the measured values ​​is denoted as , and n is the number of samples. The closer R² is to 1, the smaller the RMSE, MAE, and MAPE are, indicating that the model has better predictive performance.

[0032] Evaluation results show that the BTO-VMD-CNN-BiLSTM-Attention model outperforms the BTO-EMD-CNN-BiLSTM-Attention model in predicting land surface temperature and evaporation at various sites. In short-term predictions (1 step), the R² of the BTO-VMD model generally exceeds 95%, and the RMSE is as low as 0.35℃ for land surface temperature prediction and 0.06mm for evaporation prediction. Even in long-term prediction tasks of 90 days and 180 days, the R² of the BTO-VMD model remains above 87%, and the MAPE is as low as 2.90%, with no significant phase lag or amplitude decay. In contrast, the R² of the BTO-EMD model drops below 80% and the MAPE exceeds 6% in long-term predictions, with significant deviations from the measured values ​​in some periods, indicating poor stability. This verifies that VMD decomposition is more stable than EMD decomposition, and the BTO-VMD fusion model has better long-term prediction capabilities and robustness.

[0033] S3.5: Explanation of the Mechanism of Action of Meteorological Factors Based on the attention mechanism weight distribution of the BTO-VMD-CNN-BiLSTM-Attention model (optimal fusion model), this study analyzes the influence weights of different meteorological factors on the prediction results of land surface temperature and evaporation, revealing the complex nonlinear regulatory relationship among meteorological factors. (1) Surface temperature prediction: Air temperature (average weight 34.7%) and sunshine duration (average weight 23.1%) are the most important driving factors. Air temperature directly determines the basis of surface heat accumulation, and sunshine duration affects the amount of solar radiation received by the surface, thereby regulating the change of surface temperature. The addition of evaporation (average weight 18.6%) significantly improves the model's ability to model the surface heat balance process, reflecting the important significance of water-heat coupling in the climate system. Relative humidity (average weight 10.8%), air pressure (average weight 8.3%) and precipitation (average weight 4.5%) have relatively weak effects on surface temperature, only playing a certain regulatory role (e.g., precipitation will reduce surface temperature, and high humidity will inhibit surface heat dissipation). (2) Evaporation prediction: Air temperature (average weight 31.8%), surface temperature (average weight 25.6%), and sunshine duration (average weight 19.4%) are the three core factors of evaporation variation, which together determine the potential of surface energy for water conversion; surface temperature, as a direct reflection of surface heat accumulation, has a more significant impact on evaporation during the high evaporation period in summer; relative humidity (average weight 13.2%) has a significant inhibitory effect on evaporation rate, and the more humid the air, the more difficult it is for water to evaporate; air pressure (average weight 5.7%) and precipitation (average weight 4.3%) have the least impact; the weight distribution of different meteorological factors in the prediction of surface temperature and evaporation is shown in the table below: The above attribution analysis not only reveals the differentiated mechanisms by which major meteorological factors affect the prediction of surface temperature and evaporation, but also provides theoretical support for variable selection and physical modeling in subsequent prediction models, further enhancing the interpretability of the models.

[0034] Example 2 like Figures 1-7 As shown, this embodiment provides a long-term prediction system for land surface temperature and evaporation to implement the prediction method described in Embodiment 1. The system adopts a modular design, has a clear structure, strong scalability, and can adapt to prediction needs across multiple sites and durations. Specifically, it includes the following modules: 2.1 Data Collection and Preprocessing Module Function: Used to collect meteorological data from multiple stations, perform hierarchical preprocessing on the raw data, and output a high-quality standardized dataset; Specific components include: a data acquisition unit, a missing value imputation unit, an outlier removal unit, a data standardization unit, and a dataset partitioning unit; (1) Data acquisition unit: Collects surface temperature, evaporation and various auxiliary meteorological factors by calling the meteorological data platform API or importing local meteorological observation data. It supports parallel acquisition of data from multiple stations and supports common formats such as CSV and Excel. (2) Missing value imputation unit: A hybrid method of linear interpolation and nearest neighbor mean imputation is used to impute missing data with different missing rates to ensure the rationality of the imputed data; (3) Outlier removal unit: Outliers are identified based on the 3σ criterion and replaced by linear interpolation to avoid interference from outliers; (4) Data standardization unit: The Min-Max standardization method is used to eliminate the difference in the dimensions of variables and normalize the data to the [0,1] interval; (5) Data set partitioning unit: Divide the training set, validation set and test set according to the preset ratio, and construct a multi-step prediction dataset and output it to the model building module.

[0035] 2.2 Hybrid Prediction Model Module Function: Used to construct the improved CNN-BiLSTM-Attention hybrid prediction model described in Example 1, to achieve accurate extraction and temporal modeling of multi-source meteorological features; Specific components include: multi-scale convolutional feature extraction unit, adaptive feature calibration unit, BiLSTM temporal modeling unit, and joint weighted attention unit; (1) Multi-scale convolutional feature extraction unit: Three sets of convolutional kernels of different sizes are set to extract meteorological fluctuation features at different time scales in parallel. High-frequency noise is suppressed through feature fusion and batch normalization. (2) Adaptive feature calibration unit: calculates the mean and variance of features at each site, constructs adaptive calibration coefficients, and performs dynamic calibration of cross-site features to improve the model's generalization ability; (3) BiLSTM time series modeling unit: composed of forward LSTM and backward LSTM, which fully characterizes the time series correlation of meteorological sequences; (4) Joint weighted attention unit: realizes joint weighting of time dimension and variable / modal dimension, accurately integrates key features, and improves model interpretability.

[0036] 2.3 Modal Decomposition and Parameter Optimization Module Function: Used to perform mode decomposition on the original predicted target sequence and to achieve adaptive optimization of key parameters of the fusion model through the BTO algorithm; Specific components include: EMD decomposition unit, VMD decomposition unit, component screening unit, and BTO parameter optimization unit; (1) EMD decomposition unit: The original sequence is adaptively decomposed using the EMD method to extract IMF components and residual terms and suppress endpoint effects; (2) VMD decomposition unit: The VMD method is used to stably decompose and denoise the original sequence to avoid mode aliasing; (3) Component filtering unit: Based on the prediction loss of each IMF component, invalid components are removed and valid components are retained for subsequent modeling; (4) BTO parameter optimization unit: Construct joint parameter vector, set BTO algorithm parameters, and output the optimal parameter combination through iterative optimization, and update the mode decomposition parameters and model hyperparameters respectively.

[0037] 2.4 Long-term forecasting and results output module Function: Used for long-term prediction of surface temperature and evaporation, result reconstruction, performance evaluation and mechanism interpretation, outputting prediction results and analysis reports; Specific components include: a multi-step prediction unit, a result reconstruction unit, a performance evaluation unit, a mechanism analysis unit, and a result visualization unit; (1) Multi-step prediction unit: The selected modal components are input into the optimized fusion model to achieve multi-step prediction for the next 30 days, 90 days and 180 days; (2) Result reconstruction unit: The linear superposition method is used to reconstruct the prediction results of each modal component into the predicted values ​​of the actual physical quantities; (3) Performance evaluation unit: Calculate four evaluation indicators: R², RMSE, MAE, and MAPE, and compare the prediction performance of the two fusion models; (4) Mechanism of Action Analysis Unit: Based on the weight distribution of attention mechanism, analyze the influence weight of each meteorological factor and reveal the regulatory relationship of meteorological factors; (5) Results visualization unit: The predicted values ​​and measured values, various evaluation indicators, and the weight distribution of meteorological factors are displayed in the form of charts and graphs, and a prediction report is output, supporting data export.

[0038] 2.5 System Control Module Function: As the core control unit of the system, it coordinates the workflow of each module, receives inputs and outputs from each module, and realizes the orderly transmission and processing of data; at the same time, it supports user interaction, allowing users to set prediction parameters (such as prediction duration, number of decomposed modes, BTO algorithm parameters, etc.), view processing progress and final results, and ensure stable system operation.

[0039] Example 3 This embodiment uses meteorological data from Pingluo Station, Shizuishan Station, and Yinchuan Station as examples to run the forecasting system described in Embodiment 2. The specific operation process is as follows: 1. Start the system and collect 3,652 sets of meteorological data, including surface temperature, evaporation, and air temperature, from three stations through the data collection and preprocessing module. Complete missing value imputation, outlier removal, and standardization processing, and divide the data into training set, validation set, and test set in a 7:2:1 ratio. 2. Construct an improved CNN-BiLSTM-Attention model using the improved prediction model building module, and set the initial parameters; 3. Using the mode decomposition and parameter optimization module, the surface temperature and evaporation sequences were decomposed using EMD and VMD methods respectively, and four effective components (imf0-imf3) were selected. The BTO algorithm was then started to optimize the parameters and obtain the optimal parameter combination. 4. Through the long-term prediction and result output module, the optimal fusion model (BTO-VMD-CNN-BiLSTM-Attention) is used to predict the surface temperature and evaporation of three stations in the next 30 days, 90 days and 180 days respectively, reconstruct the prediction results, calculate the evaluation index, and analyze the mechanism of action of meteorological factors. 5. Through the results visualization unit, output a comparison chart of predicted and measured values, a bar chart of evaluation indicators, and a weight distribution chart of meteorological factors, generate a predictive analysis report, and support users to export data and reports.

[0040] The results show that the system can stably and efficiently achieve long-term prediction of surface temperature and evaporation across different sites. It has high prediction accuracy and strong robustness, and can clearly reveal the mechanism of action of meteorological factors, meeting the practical application needs of meteorological forecasting, agricultural irrigation, ecological environment assessment and other fields.

[0041] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for predicting land surface temperature and evaporation, characterized in that, Includes the following steps: Meteorological data from meteorological stations are collected, including surface temperature, evaporation, and meteorological factor sequences. A hierarchical preprocessing strategy is used to perform missing value imputation and outlier removal preprocessing on the meteorological data. A hybrid prediction model is constructed, which integrates convolutional neural networks, bidirectional long short-term memory networks, and attention mechanisms to model the nonlinear, high-dimensional, and time-series correlation features of surface temperature, evaporation, and meteorological factor sequences. The original surface temperature sequence and evaporation sequence were decomposed into multiple intrinsic mode function components of different frequencies using the mode decomposition method. The barrel theory optimization algorithm is introduced to adaptively and jointly optimize the mode decomposition parameters and the hyperparameters of the hybrid prediction model, so as to obtain the optimized fusion prediction model. The fusion prediction model is used to make multi-step predictions of surface temperature and evaporation to obtain predicted values ​​within a preset time range.

2. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, The construction of the hybrid prediction model includes: Local features are extracted from the original meteorological factor sequence using a convolutional neural network to identify local fluctuation patterns, short-term trends, and correlations in the input data. By using a bidirectional long short-term memory network to perform time-series modeling on the feature sequences extracted by the convolutional neural network, the long-term dependence of meteorological factors over time and bidirectional dynamic information can be captured. By calculating the weight distribution at different time steps through the attention layer, the system focuses on the part of the historical sequence that has the greatest influence on the current prediction, and generates a weighted global feature representation.

3. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, The mode decomposition method includes two approaches: empirical mode decomposition and variational mode decomposition. Both empirical mode decomposition and variational mode decomposition perform multi-scale decomposition on the surface temperature series and evaporation series, extracting intrinsic mode function components of different frequencies. By comparing and analyzing the decomposition stability, anti-mode aliasing ability, and subsequent prediction accuracy of the two methods, the method with higher decomposition stability and better prediction performance is selected for subsequent prediction.

4. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, The barrel theory optimization algorithm uses the mode decomposition parameters and the hyperparameters of the hybrid prediction model as unified variables to be optimized, and achieves a dynamic balance between global search and local convergence through barrel adjustment and perturbation search mechanisms.

5. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, The multi-step forecast includes predictions of surface temperature and evaporation over the next 30, 90, and 180 days.

6. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, It also includes statistical analysis of the prediction loss of each intrinsic mode function component after mode decomposition, and retaining the first few order mode components with prediction errors below a preset threshold for hybrid prediction modeling.

7. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, It also includes weight distribution analysis based on attention mechanism to analyze the influence weights of different meteorological factors on the prediction results of surface temperature and evaporation, revealing the nonlinear regulatory relationship between meteorological factors.

8. The method for predicting surface temperature and evaporation as described in claim 7, characterized in that, In the surface temperature prediction, the influence weights of each meteorological factor, from highest to lowest, are: air temperature, sunshine duration, evaporation, relative humidity, air pressure, and precipitation. In the evaporation forecast, the influence weights of each meteorological factor, from highest to lowest, are: air temperature, surface temperature, sunshine duration, relative humidity, air pressure, and precipitation.

9. The method for predicting surface temperature and evaporation as described in claim 1, characterized in that, The method supports simultaneous prediction of time-series data from multiple monitoring stations and multiple meteorological elements.

10. A surface temperature and evaporation prediction system based on the method of any one of claims 1-9, characterized in that, include: The data collection and preprocessing module is used to collect meteorological data from multiple stations and perform missing value imputation and outlier removal. The hybrid prediction model module includes a convolutional neural network layer, a bidirectional long short-term memory network layer, and an attention mechanism layer, which are used to model and predict each modal component separately. The mode decomposition module is used to decompose the original surface temperature sequence and evaporation sequence into multiple intrinsic mode function components; The optimization module is used to introduce the barrel theory optimization algorithm to adaptively and jointly optimize the mode decomposition parameters and the hyperparameters of the hybrid prediction model. The prediction execution module is used to perform multi-step predictions of surface temperature and evaporation using the optimized fusion prediction model. The weighting analysis module is used to analyze the influence weights of different meteorological factors on the prediction results based on the attention mechanism.

Images