Evaporation amount prediction method and device, storage medium and electronic equipment

Abstract

This application discloses an evaporation prediction method, apparatus, storage medium, and electronic device, relating to the field of data processing technology. The method includes: converting and correcting observation data from heterogeneous evaporating pans to construct a continuous and standardized target evaporation sequence; identifying the influence weights of target meteorological factors on evaporation based on the correlation between the target evaporation sequence and meteorological factors; selecting dominant factors from the target meteorological factors based on these influence weights and generating a weighted feature sequence; performing mode decomposition on the target evaporation sequence to extract the main prediction mode that reflects the core temporal characteristics; inputting the weighted feature sequence and the main prediction mode into a trained fusion prediction model; and outputting the evaporation prediction result after parallel multi-channel data alignment and tensor construction, cross-channel feature deep fusion, and gated temporal dependency modeling. This application can improve the accuracy and stability of evaporation prediction.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, storage medium and electronic device for predicting evaporation. Background Technology

[0002] Evaporation, as a key loss item in the hydrological cycle, is directly related to the safety of water resource allocation, the economic efficiency of water conservancy project operation, and the scientific nature of agricultural irrigation management. It is one of the core research topics in the field of hydrometeorology and a crucial technical support for the informatization of major water conservancy projects and the optimal allocation of water resources in arid and semi-arid regions. Against the backdrop of global climate change, extreme weather events are becoming more frequent, and the spatiotemporal variation characteristics of evaporation are becoming increasingly complex, placing higher demands on its prediction accuracy and stability. Accurate evaporation prediction results can provide a scientific basis for decisions on inter-regional water transfer, reservoir regulation, and water-saving irrigation in farmland, effectively reducing water waste and the operational risks of water conservancy projects.

[0003] Currently, evaporation forecasting mainly relies on statistical models based on observational data, traditional machine learning models, and single deep learning models. Among them, statistical models such as the ARIMA model predict evaporation by mining the temporal patterns of the sequence itself, traditional machine learning models such as BP neural networks and support vector machines rely on the statistical correlation between meteorological factors and evaporation to construct the prediction relationship, and single deep learning models such as Long Short-Term Memory (LSTM) networks improve the prediction effect by capturing the long-term dependencies of the sequence. However, existing technical solutions generally face core technical bottlenecks: evaporation observation data suffers from heterogeneity and discontinuity. Due to differences in structural design and thermodynamic response mechanisms, the observation data of the widely used E601 type and 20cm type evaporating pans in my country cannot be directly reused equivalently. Traditional simple ratio or linear conversion methods ignore the dynamic influence of meteorological factors, resulting in large data conversion errors and making it difficult to form a continuous and standardized training dataset. Simultaneously, evaporation is affected by the nonlinear coupling of multiple meteorological factors such as temperature, humidity, and wind speed. Existing models mostly use equal-weighted feature input methods, which cannot effectively identify dominant and weakly correlated factors, resulting in redundant input features, large noise interference during model training, and insufficient generalization ability. Furthermore, evaporation time series exhibit strong seasonality, multiple frequency components, and significant non-stationarity. Traditional models and single deep learning models struggle to effectively separate high-frequency noise and low-frequency trends in the sequence, and lack sensitivity to abrupt changes, making it difficult to meet practical engineering needs in terms of prediction accuracy and stability. Summary of the Invention

[0004] In view of this, this application provides an evaporation prediction method, apparatus, storage medium, and electronic device, which can improve the accuracy and stability of evaporation prediction.

[0005] According to a first aspect of this application, an evaporation prediction method is provided, comprising: The observation data of heterogeneous evaporating dishes were converted and corrected to construct a continuous and standardized target evaporation sequence. Based on the correlation between the target evaporation sequence and meteorological factors, the influence weight of the target meteorological factors on evaporation is identified, and the dominant factor is screened from the target meteorological factors based on the influence weight and a weighted feature sequence is generated. The target evaporation sequence is subjected to mode decomposition to extract the main prediction mode that reflects the core temporal characteristics; The fusion prediction model, trained by the weighted feature sequence and the main prediction modality input, outputs the evaporation prediction result after parallel multi-channel data alignment and tensor construction, cross-channel feature deep fusion, and gated temporal dependency modeling.

[0006] According to a second aspect of this application, an evaporation prediction device is provided, comprising: The module is used to convert and correct the observation data of heterogeneous evaporating dishes and construct a continuous and standardized target evaporation sequence. The screening module is used to identify the influence weight of the target meteorological factors on evaporation based on the correlation between the target evaporation sequence and meteorological factors, and to screen out the dominant factors from the target meteorological factors based on the influence weight and generate a weighted feature sequence. The extraction module is used to perform mode decomposition on the target evaporation sequence and extract the main prediction mode that reflects the core time series characteristics; The input module is used to input the weighted feature sequence and the fusion prediction model trained by the main prediction modality into the fusion prediction model, and output the evaporation prediction result after parallel multi-channel data alignment and tensor construction, cross-channel feature deep fusion and gated temporal dependency modeling.

[0007] According to a third aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described evaporation prediction method.

[0008] According to a fourth aspect of this application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described evaporation prediction method.

[0009] By employing the aforementioned technical solutions, the evaporation prediction method, apparatus, storage medium, and electronic equipment provided in this application, through conversion and correction of heterogeneous evaporation pan observation data to construct a continuous standardized target evaporation sequence, can effectively solve the problems in existing technologies where heterogeneous evaporation data cannot be directly reused and where large errors in traditional conversion methods lead to insufficient training dataset quality. By identifying the influence weights of target meteorological factors on evaporation and selecting dominant factors to generate a weighted feature sequence, redundant interference and noise problems caused by equally weighted feature inputs can be avoided, improving the model's generalization ability. By performing modal decomposition on the target evaporation sequence to extract the master prediction mode of core temporal features, effective decomposition of multi-frequency components and seasonal characteristics in the evaporation time series can be achieved, enhancing the sensitivity to sequence abrupt change points. Finally, by combining the parallel multi-channel data processing, cross-channel feature fusion, and gated temporal dependency modeling of the fusion prediction model, the complementary advantages of dominant factor features and core temporal features can be fully integrated, significantly improving the accuracy and stability of evaporation prediction and meeting the needs of practical engineering applications.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an evaporation prediction method provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating an evaporation prediction method according to another embodiment of this application is shown; Figure 3 This paper shows a schematic diagram of the structure of an evaporation prediction device provided in an embodiment of this application; Figure 4 A schematic diagram of an evaporation prediction device provided in another embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] Currently, evaporation forecasting mainly relies on statistical models based on observational data, traditional machine learning models, and single deep learning models. Among them, statistical models such as the ARIMA model predict evaporation by mining the temporal patterns of the sequence itself, traditional machine learning models such as BP neural networks and support vector machines rely on the statistical correlation between meteorological factors and evaporation to construct the prediction relationship, and single deep learning models such as Long Short-Term Memory (LSTM) networks improve the prediction effect by capturing the long-term dependencies of the sequence. However, existing technical solutions generally face core technical bottlenecks: evaporation observation data suffers from heterogeneity and discontinuity. Due to differences in structural design and thermodynamic response mechanisms, the observation data of the widely used E601 type and 20cm type evaporating pans in my country cannot be directly reused equivalently. Traditional simple ratio or linear conversion methods ignore the dynamic influence of meteorological factors, resulting in large data conversion errors and making it difficult to form a continuous and standardized training dataset. Simultaneously, evaporation is affected by the nonlinear coupling of multiple meteorological factors such as temperature, humidity, and wind speed. Existing models mostly use equal-weighted feature input methods, which cannot effectively identify dominant and weakly correlated factors, resulting in redundant input features, large noise interference during model training, and insufficient generalization ability. Furthermore, evaporation time series exhibit strong seasonality, multiple frequency components, and significant non-stationarity. Traditional models and single deep learning models struggle to effectively separate high-frequency noise and low-frequency trends in the sequence, and lack sensitivity to abrupt changes, making it difficult to meet practical engineering needs in terms of prediction accuracy and stability.

[0014] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting evaporation, such as... Figure 1 As shown, the method includes: Step 110: Convert and correct the observation data of heterogeneous evaporating dishes to construct a continuous and standardized target evaporation sequence.

[0015] Among them, heterogeneous evaporation dish observation data refers to evaporation observation data collected from evaporation dishes of different types, structural designs, or thermodynamic response mechanisms (such as E601 type evaporation dish and 20cm diameter evaporation dish). Due to the inherent differences in observation equipment, this type of data has systematic bias and cannot be directly reused equivalently. Conversion correction refers to the data processing process of eliminating systematic bias between heterogeneous evaporation dish observation data through specific technical means and unifying heterogeneous data from different sources under the same standard. Continuous standardized target evaporation sequence refers to the evaporation data sequence formed after conversion correction and data integration, which has continuous time dimension, unified data standard, and controllable bias, and can be directly used for subsequent model training, feature extraction, and predictive analysis.

[0016] In this embodiment of the disclosure, to address the inherent biases and time discontinuities in heterogeneous observation data collected from different types of evaporating dishes, auxiliary influencing variables that are significantly correlated with the heterogeneous data bias can be screened first. Then, an appropriate conversion model (such as a multiple regression model, a machine learning correction model, etc.) is used to adjust the model parameters with the auxiliary variables as the core, and the heterogeneous evaporating dish observation data is uniformly converted and corrected to eliminate systematic errors caused by different devices. Finally, through data completion, alignment, and other integration processing in the time dimension, a target evaporation sequence with continuous time dimension and unified data standards is constructed, providing a high-quality data foundation for subsequent evaporation prediction-related analysis.

[0017] This technical step effectively solves the problems in existing technologies where observation data from heterogeneous evaporation pans cannot be directly and equivalently reused, and where traditional conversion methods ignore dynamic effects, leading to large errors. By constructing a continuous and standardized target evaporation sequence, it can not only eliminate the heterogeneity and discontinuity of the data, significantly improving the consistency and usability of the data, but also provide reliable data support for subsequent meteorological factor weight identification, temporal feature extraction, and fusion prediction. This ensures the training quality of the evaporation prediction model from the source and lays a solid foundation for improving the overall prediction accuracy and stability.

[0018] Step 120: Based on the correlation between the target evaporation sequence and meteorological factors, identify the influence weight of the target meteorological factors on evaporation, and select the dominant factors from the target meteorological factors based on the influence weights and generate a weighted feature sequence.

[0019] Among them, the target meteorological factors refer to various meteorological elements that have potential correlation with changes in evaporation and are used to analyze their impact on evaporation after preliminary screening. These include, but are not limited to, air temperature, relative humidity, wind speed, sunshine duration, air pressure, surface temperature, and solar radiation. The influence weight is a numerical indicator used to quantify the degree of contribution or magnitude of influence of each target meteorological factor to changes in evaporation. The higher the weight value, the stronger the driving effect of the corresponding meteorological factor on evaporation. The dominant factor is a key meteorological factor selected from the target meteorological factors that has reached the preset standard in influence weight and plays a core driving role in changes in evaporation. It is different from weakly correlated factors that have a weak impact on evaporation. The weighted feature sequence is a feature sequence that can accurately reflect the dynamic impact of meteorological factors on evaporation by integrating the weighted data of all dominant factors along the time dimension after weighting the original observation data of the dominant factor with its corresponding influence weight.

[0020] In this embodiment of the disclosure, based on the constructed continuous standardized target evaporation sequence, statistical analysis, machine learning, or other suitable quantitative analysis methods can be used to explore the intrinsic correlation between the target evaporation sequence and various meteorological factors. Then, through quantitative modeling, the influence weight of each target meteorological factor on the change of evaporation can be identified. Subsequently, according to the preset screening rules, the dominant factors that play a major driving role in evaporation are screened from the target meteorological factors. The original observation data of each dominant factor and its corresponding influence weight are then weighted and fused. The weighted data of all dominant factors are integrated according to the time dimension, and finally a weighted feature sequence that can highlight the core meteorological influence is formed, providing accurate feature input for subsequent evaporation prediction.

[0021] This technical step quantifies and identifies the influence weights of meteorological factors on evaporation, enabling precise differentiation of the degree of influence of meteorological factors. The process of screening dominant factors can effectively eliminate redundant interference from weakly correlated factors, reduce noise input in subsequent model training, and significantly improve the effectiveness of feature input. At the same time, the generated weighted feature sequence can strengthen the core driving role of dominant factors on evaporation, accurately characterize the correlation between meteorological factors and evaporation, not only reducing the computational complexity of the model, but also providing high-quality feature support for subsequent prediction models, effectively improving the model's generalization ability and prediction accuracy.

[0022] Step 130: Perform mode decomposition on the target evaporation sequence and extract the main prediction mode that reflects the core time series characteristics.

[0023] Mode decomposition refers to the process of breaking down a complex non-stationary time series into multiple independent modal components with different frequency characteristics and relatively stable operation through a specific algorithm. Each modal component corresponds to an inherent variation law of the original series. Core time series features refer to the inherent laws that play a dominant role in the variation of evaporation, including but not limited to key time series characteristics such as seasonal fluctuations, long-term trends, and periodic changes. These are the core elements that determine the variation of evaporation. The main prediction mode refers to the set of modal components selected from the multiple independent modal components obtained from mode decomposition that can best reflect the core time series features of evaporation and have a key supporting role in the prediction results.

[0024] In this embodiment of the disclosure, to address the non-stationarity and multi-frequency characteristics of the target evaporation sequence, a mode decomposition algorithm (such as variational mode decomposition, empirical mode decomposition, etc.) can be used to decompose the sequence into multiple independent mode components with different frequency characteristics. Subsequently, each mode component is quantitatively evaluated through a multi-dimensional evaluation system (such as energy ratio, stationarity test, correlation analysis, etc.) to select mode components that can accurately characterize the core temporal pattern of evaporation. Finally, the selected mode components are integrated to form the main prediction mode for subsequent prediction analysis, providing high-quality temporal feature input for the fusion prediction model.

[0025] This technical step decomposes complex non-stationary evaporation sequences into simple, stationary modal components through modal decomposition. This effectively separates high-frequency noise from the core temporal patterns in the sequence, overcoming the technical bottleneck of traditional models in handling multi-frequency, non-stationary time-series data. Simultaneously, by selecting and extracting the main prediction mode, it can accurately focus on the core temporal features that dominate evaporation changes, eliminating interference from redundant noise components. This not only simplifies the temporal modeling difficulty of subsequent models but also provides accurate temporal feature support for the fusion prediction model, significantly improving the model's ability to capture evaporation change patterns and laying a solid foundation for improving overall prediction accuracy and stability.

[0026] Step 140: The fusion prediction model trained by the weighted feature sequence and the main prediction modality is then used for parallel multi-channel data alignment and tensor construction, deep fusion of cross-channel features, and gated temporal dependency modeling to output the evaporation prediction result.

[0027] Among them, the fusion prediction model refers to a prediction model with the ability to integrate multi-source features and time series modeling, capable of simultaneously processing meteorological features and time series features; parallel multi-channel data alignment refers to using a multi-channel parallel processing method to synchronously match the weighted feature sequence with the main prediction mode in the time dimension, ensuring that the timestamps of the two types of data are consistent; tensor construction refers to converting the aligned weighted feature sequence with the main prediction mode in data format to construct a tensor data structure that adapts to the input requirements of the fusion prediction model; cross-channel feature deep fusion refers to breaking down the channel barriers between the weighted features and the main prediction mode through the model's specific fusion mechanism, deeply integrating the meteorological factor correlation features with the evaporation time series features, and generating fusion features that can comprehensively reflect the two types of core information; gated time series dependency modeling refers to the fusion prediction model capturing the long-term and short-term variation correlation of evaporation in the fusion features through gating mechanisms (such as gated cyclic units, long short-term memory unit gating structures, etc.), achieving accurate modeling of time series patterns.

[0028] In this embodiment of the disclosure, the weighted feature sequence that can characterize the core meteorological impact and the main prediction mode that reflects the core temporal pattern of evaporation can be used as multi-source input data. The two types of data are synchronized in the time dimension through parallel multi-channel processing. Then, the aligned two types of data are converted into tensor format that is suitable for model input. Subsequently, the cross-channel fusion mechanism of the fusion prediction model is used to deeply integrate the tensorized meteorological features and temporal features to generate fusion features that contain both types of core information. Finally, the long-term and short-term variation patterns of evaporation in the fusion features are accurately captured through the model's built-in gated temporal dependency modeling mechanism. After model calculation, the final evaporation prediction result is output.

[0029] This technical process ensures the temporal consistency of multi-source input data through parallel multi-channel data alignment, avoiding prediction bias caused by data asynchrony. Deep fusion of cross-channel features enables the complementary advantages of meteorological factor correlation features and evaporation time-series features, overcoming the limitations of single feature types on prediction performance. Gated time-series dependency modeling accurately captures the long-term and short-term variation patterns of evaporation, effectively solving the problem of traditional models struggling to handle complex time-series dependencies. Finally, through the full-process processing of the fusion prediction model, multi-source core information is fully integrated, significantly improving the accuracy and stability of evaporation prediction.

[0030] In summary, the evaporation prediction method provided in this application effectively solves the problems of existing technologies, such as the inability to directly reuse heterogeneous evaporation data and the large errors in traditional conversion methods leading to insufficient training dataset quality, by converting and correcting the observation data of heterogeneous evaporation pans to construct a continuous standardized target evaporation sequence. By identifying the influence weights of target meteorological factors on evaporation and selecting dominant factors to generate a weighted feature sequence, redundant interference and noise problems caused by equal-weighted feature inputs can be avoided, improving the model's generalization ability. By performing modal decomposition on the target evaporation sequence to extract the master prediction mode of core temporal features, the multi-frequency components and seasonal characteristics in the evaporation time series can be effectively separated, enhancing the sensitivity to sequence abrupt change points. Finally, by combining the parallel multi-channel data processing, cross-channel feature fusion, and gated temporal dependency modeling of the fusion prediction model, the complementary advantages of dominant factor features and core temporal features can be fully integrated, significantly improving the accuracy and stability of evaporation prediction and meeting the needs of practical engineering applications.

[0031] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the implementation of this embodiment, this embodiment also provides another method for predicting evaporation, such as... Figure 2 As shown, the method includes: Step 210: Select meteorological factors that are significantly correlated with the differences in the observation data of heterogeneous evaporation pans as adjustment variables, and construct a multiple regression conversion model that includes the adjustment variables.

[0032] Among them, the meteorological factors with significant correlations refer to meteorological elements (such as temperature, humidity, and wind speed) that, through quantitative analysis, have a strong correlation with the differences in the observation data of the heterogeneous evaporating pans and can dynamically affect the magnitude of the differences. The adjustment variables are the core variables selected from the meteorological factors with significant correlations and used to be incorporated into the conversion model to specifically correct the differences in the heterogeneous evaporating pan data. Their role is to reduce data bias by quantifying the influence of meteorological factors. The multiple regression conversion model is a mathematical model based on multiple regression analysis. It takes the selected adjustment variables as input and aims to correct the differences in the heterogeneous evaporating pan data. By establishing a quantitative correlation between variables and data differences, it achieves accurate conversion of heterogeneous data.

[0033] In this embodiment of the disclosure, meteorological factors that are significantly associated with the differences in observation data of heterogeneous evaporating pans can be screened first through quantitative means such as correlation analysis. These meteorological factors are identified as adjustment variables that can dynamically correct the differences in data. Then, based on the multiple regression analysis method, a multiple regression conversion model is constructed with these adjustment variables as the core input and can quantitatively characterize the relationship between the adjustment variables and the differences in data, so as to provide model support for the accurate correction of subsequent observation data of heterogeneous evaporating pans.

[0034] This technical step, by selecting meteorological factors that are significantly correlated with the differences in heterogeneous data as adjustment variables, ensures the targeted nature of model calibration and avoids the drawbacks of traditional conversion methods that ignore the dynamic influence of meteorological events. Furthermore, the multivariate regression conversion model that includes adjustment variables can quantify and integrate the comprehensive influence of multiple meteorological factors, effectively reducing the systematic bias of heterogeneous evaporation pan observation data, improving the accuracy of data conversion, and laying a reliable model foundation for the subsequent construction of a continuous and standardized target evaporation sequence, thereby improving the quality of data processing related to evaporation prediction from the source.

[0035] Step 220: Based on the multivariate regression conversion model, the observation data of heterogeneous evaporating dishes are converted and corrected, and the converted and corrected observation data of heterogeneous evaporating dishes are integrated to obtain a continuous standardized target evaporation sequence.

[0036] In this embodiment of the disclosure, a fully constructed multivariate regression conversion model can be used as the core tool to input heterogeneous observation data collected from different types of evaporation pans into the model for deviation correction, so that all types of data are unified to the same standard. Subsequently, all corrected data are sorted and integrated in the time dimension to complete data alignment, completion and other processing, eliminating data discontinuity and fragmentation problems, and finally forming a time-continuous and standardized target evaporation sequence.

[0037] This technical step, through a multivariate regression conversion model, enables precise correction of heterogeneous evaporation pan data. It effectively solves the problem of large errors caused by traditional conversion methods ignoring the influence of meteorological dynamics, and can significantly reduce the systematic bias of the data. At the same time, by integrating and processing the corrected data, it can eliminate data discontinuities. The constructed continuous standardized target evaporation sequence can significantly improve the consistency and usability of the data. It can not only provide high-quality data support for subsequent meteorological factor weight identification and time series feature extraction, but also ensure the training quality of subsequent prediction models from the source, laying a solid foundation for improving the overall accuracy and stability of evaporation prediction.

[0038] Step 230: Based on the correlation between the target evaporation sequence and meteorological factors, identify the influence weight of the target meteorological factors on evaporation.

[0039] For embodiments of this disclosure, step 230 may include the following steps: Step 230-1: Perform correlation analysis between the target evaporation sequence and multiple preset meteorological factors, and screen multiple target meteorological factors whose correlation is greater than the preset correlation threshold.

[0040] Among them, the preset meteorological factors refer to various meteorological elements that are pre-selected and may be related to changes in evaporation, including but not limited to core meteorological parameters such as temperature, relative humidity, wind speed, sunshine duration, air pressure, surface temperature, and solar radiation; correlation analysis refers to the data analysis process that measures the strength of the correlation between the target evaporation sequence and each preset meteorological factor through quantitative analysis methods (such as Pearson correlation coefficient, Spearman correlation coefficient, etc.); the preset correlation threshold refers to the pre-set quantitative standard used to judge whether the correlation is effective, which is the key basis for screening core correlated meteorological factors. If the correlation value is higher than the threshold, it is considered that there is a significant correlation between the two.

[0041] In this embodiment of the disclosure, a continuous and standardized target evaporation sequence can be used as the analysis benchmark. Multiple preset meteorological factors, including temperature and relative humidity, are selected. The correlation between the target evaporation sequence and each preset meteorological factor is analyzed using an appropriate quantitative analysis method to obtain the quantitative correlation results between each preset meteorological factor and evaporation. Then, each quantitative result is compared with a preset correlation threshold, and meteorological factors with correlation values ​​higher than the threshold are selected and identified as target meteorological factors that have a significant impact on evaporation.

[0042] This technical step, through correlation analysis, can accurately quantify the degree of correlation between meteorological factors and evaporation. By using a preset correlation threshold for screening, redundant meteorological factors that are weakly or uncorrelated with evaporation can be effectively eliminated, reducing noise interference in subsequent analyses. At the same time, it identifies target meteorological factors that have a significant impact on evaporation, laying a precise foundation for accurately identifying the influence weight of each meteorological factor on evaporation and screening dominant factors. This not only simplifies the complexity of subsequent analyses but also improves the pertinence and effectiveness of the correlation analysis of meteorological factors in the entire technical solution, providing key support for ultimately improving the accuracy of evaporation prediction.

[0043] Step 230-2: Using the target evaporation sequence as the output variable and the target meteorological factors as the input variables, establish a support vector machine regression model based on radial basis kernel function.

[0044] Among them, the radial basis function (RBF) is a commonly used kernel function that can map input variables to a high-dimensional feature space, effectively handling the nonlinear correlation between input and output, and providing support for the support vector machine model to capture complex mapping relationships. The support vector machine regression model (SVM) is a regression prediction model built based on the support vector machine algorithm. By finding the optimal hyperplane to fit the correlation between the input variable (target meteorological factor) and the output variable (target evaporation sequence), it has strong nonlinear fitting ability and generalization performance.

[0045] In this embodiment of the disclosure, a target meteorological factor that has a significant impact on evaporation can be used as the model input, and a target evaporation sequence that characterizes the change in evaporation can be used as the model output. A radial basis kernel function adapted to nonlinear correlation modeling is selected as the core mapping tool, and a regression model is constructed based on the support vector machine algorithm framework. Through this model, a quantitative mapping relationship between the target meteorological factor and the target evaporation sequence is constructed, providing a basic model architecture for subsequent model optimization and calculation of the influence weight of meteorological factors.

[0046] This technical step, by leveraging the nonlinear mapping capability of the radial basis kernel function, can effectively overcome the limitation of traditional linear models in fitting the complex nonlinear relationship between meteorological factors and evaporation. The constructed support vector machine regression model has strong fitting accuracy and generalization ability, and can accurately quantify the intrinsic relationship between target meteorological factors and evaporation. It can provide reliable model support for subsequent model parameter optimization and accurate calculation of the influence weight of meteorological factors, significantly improving the scientificity and accuracy of the analysis of the influence of meteorological factors on evaporation.

[0047] Step 230-3: Optimize the kernel parameters and penalty factor of the support vector machine regression model, and use the optimized support vector machine model to train and cross-validate each target meteorological factor.

[0048] Among them, kernel parameters specifically refer to the key parameters of the radial basis kernel function in the support vector machine regression model (such as the bandwidth parameter of the Gaussian kernel). Their values ​​directly affect the high-dimensional mapping effect of the kernel function on the input variables, and thus determine the model's ability to fit nonlinear relationships. The penalty factor is a regularization parameter in the support vector machine model, used to balance the model's fitting accuracy and generalization ability on the training data. By controlling the intensity of the penalty for training errors, it can prevent the model from reducing its adaptability to new data due to overfitting the training data. Cross-validation refers to dividing the dataset into multiple training subsets and validation subsets, and evaluating the model's performance through multiple alternating training and validation. It is used to comprehensively test the model's generalization ability and stability, and avoid the bias caused by evaluating a single dataset.

[0049] In this embodiment of the disclosure, with the goal of improving the model fitting accuracy and generalization ability, the kernel parameters and penalty factors of the support vector machine regression model are adjusted using an adaptive parameter optimization method to determine the optimal parameter combination. Then, the optimized model is applied to the dataset corresponding to each target meteorological factor. Through data-driven learning, the model learns the intrinsic relationship between the target meteorological factor and the target evaporation sequence. At the same time, cross-validation is used to alternately complete model training and performance evaluation to comprehensively test the stability and adaptability of the model.

[0050] This technical step, by optimizing kernel parameters and penalty factors, can effectively solve the problems of overfitting or underfitting that may exist in the model, and significantly improve the fitting accuracy of the support vector machine regression model for the nonlinear relationship between meteorological factors and evaporation. Combined with multiple training and evaluation through cross-validation, the generalization ability and stability of the model can be fully verified, avoiding the bias caused by single training and evaluation, and ensuring the reliability of the model. The final optimized model can provide high-quality model support for the subsequent accurate calculation of the influence weight of each target meteorological factor on evaporation, further improving the scientificity and accuracy of meteorological factor influence analysis.

[0051] Step 230-4: Using the support vector machine regression model after training and cross-validation, calculate the influence weight of each target meteorological factor on evaporation.

[0052] In this embodiment of the disclosure, a support vector machine regression model that has been trained and cross-validated and has stable and reliable performance can be used as the core tool. Based on the accurate quantitative correlation between the target meteorological factors and evaporation that the model has learned, the contribution of each target meteorological factor to the change in evaporation can be calculated through quantitative calculation methods such as coefficient analysis and sensitivity analysis built into the model, and finally the influence weight value corresponding to each target meteorological factor can be obtained.

[0053] Step 240: Based on the influence weight, screen out the dominant factors from the target meteorological factors and generate a weighted feature sequence.

[0054] For embodiments of this disclosure, step 240 may include the following steps: Step 240-1: Normalize the influence weights of each target meteorological factor to obtain standardized weights.

[0055] Normalization refers to the mathematical process of mapping influence weights of different magnitudes to a unified numerical range (usually 0-1) using specific mathematical algorithms (such as linear normalization, L2 normalization, etc.). The core purpose is to eliminate dimensional differences and achieve horizontal comparability of different weights. Standardized weights refer to the weight values ​​obtained after normalization, which are in a unified numerical range and can be directly compared horizontally. They can intuitively reflect the relative magnitude of the influence of each target meteorological factor on evaporation.

[0056] In this embodiment of the disclosure, the influence weights of each target meteorological factor can be used as the processing object. An appropriate normalization algorithm is used to perform a mathematical transformation on all influence weights on a uniform scale, mapping the original weight values ​​with different magnitudes to a preset uniform numerical range, eliminating the dimensional barriers between different weights, and finally obtaining the standardized weights corresponding to each target meteorological factor that are horizontally comparable.

[0057] This technical step transforms the influence weights of different magnitudes into standardized weights of a uniform scale through normalization processing. This effectively eliminates the dimensional differences of the original weights, ensuring fair horizontal comparability of the influence of each target meteorological factor. Consequently, it provides an objective and unified quantitative basis for subsequent screening of dominant factors based on weights, avoiding screening bias caused by inconsistent weight magnitudes.

[0058] Step 240-2: Identify the target meteorological factors whose corresponding standardized weights are greater than the preset weight screening threshold as the dominant factors.

[0059] Among them, the preset weight screening threshold refers to the pre-set quantitative standard used to define the degree of influence of meteorological factors. It is the key basis for distinguishing core influencing factors from weakly correlated factors and is used to determine whether the target meteorological factor has a dominant role. The dominant factor refers to the target meteorological factor whose standardized weight is greater than the preset weight screening threshold. These factors have a high degree of contribution to the change of evaporation and are the core meteorological elements driving the change of evaporation.

[0060] In this embodiment of the disclosure, the standardized weights corresponding to each target meteorological factor can be used as the basis for judgment. The standardized weights of each target meteorological factor are compared with the pre-set weight screening threshold one by one. Target meteorological factors whose standardized weight values ​​exceed the threshold are directly identified as the dominant factors that play a core driving role in the change of evaporation.

[0061] This technical step compares and filters standardized weights by setting a preset weight screening threshold, which can accurately identify the dominant factors, effectively eliminate weakly correlated meteorological factors with low standardized weights, and significantly reduce redundant information and noise interference in subsequent feature processing. At the same time, by identifying the core meteorological factors driving changes in evaporation, it can provide accurate core element support for the subsequent generation of weighted feature sequences, which can not only improve the targeting and efficiency of feature processing, but also reduce the training complexity of subsequent prediction models, laying the foundation for improving the model's generalization ability and prediction accuracy.

[0062] Step 240-3: Extract the original observation data sequence corresponding to each dominant factor, and perform a dot product operation between the original observation data sequence and the corresponding standardized weight to obtain the weighted data sequence of each dominant factor.

[0063] Among them, the original observation data sequence refers to the sequence formed by arranging the original meteorological observation data collected continuously within the observation period for each dominant factor without weighting or other processing, according to the time dimension. It is the basic data carrier for characterizing the dynamic changes of the dominant factor. The weighted data sequence refers to the sequence obtained after dot product operation, which integrates the original dynamic information of the dominant factor and the corresponding influence weight, and can highlight the core driving role of the dominant factor on evaporation.

[0064] In this embodiment of the disclosure, the original meteorological observation data sequence of each identified dominant factor within the corresponding observation period can be extracted first to ensure the integrity and accuracy of the time dimension of the data sequence. Then, the original observation data sequence of each dominant factor is multiplied element by element with the standardized weight corresponding to the factor. The core influence information is enhanced by the modulation effect of the weight on the original data, and finally, a weighted data sequence that integrates the weight information is obtained for each dominant factor.

[0065] This technical step extracts the original observation data sequence of the dominant factors and performs dot-multiplication and weighted operations to achieve deep integration of the original dynamic information of the dominant factors and their influence weights. This not only preserves the temporal variation characteristics of the dominant factors but also strengthens the core influence through weight modulation, effectively highlighting information that plays a key role in changes in evaporation. At the same time, the generation of weighted data sequences avoids interference from weakly correlated factors, providing high-quality core input for subsequent feature integration. This not only improves the targeting and effectiveness of subsequent feature processing but also helps the prediction model more accurately capture the correlation between meteorological factors and evaporation, laying a solid feature foundation for improving prediction accuracy.

[0066] Step 240-4: Concatenate the weighted data sequences of all dominant factors along the time dimension to form a unified weighted feature sequence.

[0067] In this embodiment of the disclosure, the weighted data sequences corresponding to all dominant factors can be used as the processing object. First, the integrity and consistency of each weighted data sequence in the time dimension are ensured. Then, the weighted data sequences of different dominant factors are synchronized and horizontally spliced ​​with the timestamp as a unified benchmark. Multiple scattered single dominant factor weighted sequences are integrated into a single sequence containing all core meteorological impact information, and finally a unified weighted feature sequence that can be directly used for subsequent prediction modeling is formed.

[0068] Step 250: Perform mode decomposition on the target evaporation sequence and extract the main prediction mode that reflects the core time series characteristics.

[0069] For embodiments of this disclosure, step 250 may include the following steps: Step 250-1: The target evaporation sequence is decomposed using the variational mode decomposition algorithm to obtain multiple stationary intrinsic mode functions (IMF) components. Each IMF component corresponds to a different frequency feature of the target evaporation sequence.

[0070] Among them, the variational mode decomposition algorithm is an adaptive mode decomposition algorithm. Its core is to construct and solve a variational model to decompose a complex non-stationary time series into multiple stationary mode components with clear frequency centers and finite bandwidths. It has the characteristics of strong resistance to noise interference and stable decomposition results. The intrinsic mode function (IMF) component refers to the basic building block obtained after variational mode decomposition. It is a single-component signal that meets the stationarity requirement. Each component has its own unique frequency characteristics, which together constitute the time series variation law of the original target evaporation series. The frequency characteristics refer to the variation law of different time scales contained in the target evaporation series, covering various time series characteristics such as high-frequency fluctuations (such as short-term random noise), mid-frequency fluctuations (such as seasonal changes), and low-frequency trends (such as long-term evolution laws).

[0071] In this embodiment of the disclosure, a continuous and standardized target evaporation sequence can be used as the processing carrier. The variational mode decomposition algorithm is used to adaptively mine the intrinsic frequency characteristics of the sequence through the built-in variational optimization mechanism of the algorithm. The originally complex non-stationary evaporation sequence is decomposed into multiple independent intrinsic mode function (IMF) components that meet the stationarity requirements. Each IMF component corresponds to a specific frequency feature in the original sequence, thereby realizing the multi-scale decomposition of the temporal pattern of evaporation.

[0072] This technical step leverages the adaptive decomposition capability of the variational mode decomposition algorithm to effectively address the challenge of traditional methods in separating the multi-frequency and non-stationary characteristics of evaporation sequences. It simplifies complex time-series data into multiple stationary IMF components, enabling precise separation of high-frequency noise from mid-to-low-frequency core time-series patterns. Simultaneously, each IMF component clearly corresponds to different frequency characteristics, providing accurate basic units for subsequent targeted screening of core time-series information and removal of redundant noise. This significantly reduces the difficulty of subsequent time-series feature extraction, thereby laying a solid foundation for improving the ability of evaporation prediction models to capture time-series patterns.

[0073] Step 250-2: Perform multidimensional evaluation on each Intrinsic Mode Function (IMF) component, including calculating the energy proportion of each IMF component, verifying stationarity using the ADF test, and analyzing the correlation with the target evaporation sequence using the Pearson correlation coefficient method.

[0074] Among them, multidimensional assessment refers to a comprehensive evaluation method that quantitatively analyzes and verifies IMF components from multiple different dimensions (energy contribution, stationarity, and correlation) to ensure the comprehensive and accurate screening of core effective components; energy proportion refers to the proportion of the energy of a single IMF component to the total energy of the target evaporation sequence, used to quantitatively characterize the contribution of the component to the time series characteristics of the original sequence, the higher the proportion, the greater the contribution; the ADF test method, also known as the unit root test method, is a commonly used algorithm for verifying the stationarity of time series. It judges the stationarity of the sequence by testing whether there is a unit root, and is a core method for evaluating the reliability of time series data; stationarity verification refers to the process of determining whether the IMF component meets the stationarity requirements through a specific algorithm (such as the ADF test). Stationarity is a prerequisite for ensuring that the time series characteristics can be used for subsequent modeling and analysis; the Pearson correlation coefficient method is a statistical method for quantitatively analyzing the degree of linear correlation between two variables. By calculating the correlation coefficient (ranging from -1 to 1), it measures the degree of correlation between the IMF component and the target evaporation sequence.

[0075] In this embodiment of the disclosure, all intrinsic mode functions (IMFs) obtained from variational mode decomposition can be used as evaluation objects. A comprehensive evaluation is carried out from three core dimensions: energy contribution, stationarity, and correlation. The energy proportion of each IMF component is calculated to quantify its characteristic contribution to the original target evaporation sequence. The stationarity of each IMF component is verified one by one using the ADF test method to ensure that it meets the basic requirements for subsequent modeling. At the same time, the linear correlation between each IMF component and the original target evaporation sequence is analyzed using the Pearson correlation coefficient method to form a comprehensive quantitative evaluation result for all IMF components.

[0076] This technical step involves a multi-dimensional evaluation of the IMF components from three dimensions: energy proportion, stability, and correlation. This enables precise identification of the effectiveness of each component, allowing for the selection of core components that contribute significantly to the temporal characteristics of evaporation, exhibit good stability, and strong correlation. It also effectively eliminates invalid components such as high-frequency noise and weak correlations, avoiding the one-sidedness of single-dimensional evaluation. Through a comprehensive quantitative evaluation process, it ensures that the subsequently selected master prediction mode accurately reflects the core temporal patterns of evaporation, providing high-quality temporal feature support for the fusion prediction model. This significantly improves the model's ability to capture the changing patterns of evaporation, thereby ensuring the overall prediction accuracy and stability.

[0077] Step 250-3: Based on the multidimensional evaluation results, select the target intrinsic mode function (IMF) components whose energy proportion is higher than the preset energy threshold, whose stationarity test results meet the preset stationarity standard, and whose correlation with the target evaporation sequence is higher than the preset correlation coefficient threshold. The target intrinsic mode function (IMF) components are determined as the main prediction modes that can reflect the core temporal characteristics of evaporation.

[0078] Among them, the multidimensional evaluation results refer to the set of data such as energy proportion values, stationarity verification conclusions, and correlation coefficients obtained after quantitatively evaluating each intrinsic mode function (IMF) component from three core dimensions: energy contribution, stationarity, and correlation. This serves as the direct basis for selecting core time-series components. The preset energy threshold is a pre-set critical value for energy proportion, used to define the minimum standard for the contribution of IMF components to the time-series characteristics of the original target evaporation sequence. The preset stationarity standard is a pre-set quantitative criterion (such as the significance level threshold in the ADF test) used to determine whether IMF components meet the stationarity requirements, and is a fundamental prerequisite for ensuring that time-series data can be used for subsequent modeling and analysis. The preset correlation... The coefficient threshold is a pre-set correlation critical value, which is used to measure the minimum standard of the correlation between the IMF component and the original target evaporation sequence, ensuring that the selected components are directly related to the change in evaporation. The target intrinsic mode function IMF component refers to the IMF component that simultaneously meets the following conditions: energy proportion higher than the preset energy threshold, stationarity test meets the preset standard, and correlation with the target evaporation sequence is higher than the preset threshold. It is the carrier of the core time series features in the original sequence. The main prediction mode is a feature set composed of all target intrinsic mode function IMF components, which can accurately reflect the core time series law of evaporation (such as seasonal fluctuations, long-term trends, etc.) and is the core time series input of the subsequent fusion prediction model.

[0079] In this embodiment of the disclosure, the multidimensional evaluation results of each Intrinsic Mode Function (IMF) component can be used as the screening basis. The energy proportion of each IMF component is compared with a preset energy threshold, the stationarity test result is compared with a preset stationarity standard, and the correlation coefficient is compared with a preset correlation coefficient threshold. IMF components that meet the above three conditions are screened out. These target IMF components that meet all screening criteria are integrated and determined as the main prediction mode that can accurately characterize the core temporal characteristics of evaporation.

[0080] Step 260: The fusion prediction model trained by the weighted feature sequence and the main prediction modality is then used for parallel multi-channel data alignment and tensor construction, deep fusion of cross-channel features, and gated temporal dependency modeling to output the evaporation prediction result.

[0081] The fusion prediction model comprises a multi-layer memory unit, a cross-channel fusion layer, and a fully connected output layer. The multi-layer memory unit (layers 1-3) is the core component responsible for capturing temporal patterns in the fusion prediction model. It possesses the ability to mine long-term and short-term temporal dependencies and can accurately extract core temporal features from the sequence. The cross-channel fusion layer is a functional layer in the model used to break down channel barriers between different types of features (meteorological features, temporal features) and achieve deep integration of multi-source features. It can generate fusion features that combine the advantages of multiple types of features. The fully connected output layer is the output component of the model, which converts the fusion features into quantitative results that meet prediction requirements (such as predicted evaporation values) through linear mapping.

[0082] In specific application scenarios, when pre-training the fusion prediction model, the implementation steps may include: selecting historical weighted feature sequence fragments and corresponding historical main prediction mode fragments as training data, simultaneously using the actual historical evaporation values ​​corresponding to each main prediction mode fragment as training labels, and dividing the training data and training labels into training sets and validation sets according to a preset ratio; inputting the training set data and corresponding training labels into the constructed fusion prediction model, extracting sample nonlinear coupling features between dominant factors based on the historical weighted feature sequence through the model input gate, capturing the core temporal regularity features of samples in the historical main prediction mode through multi-layer memory units, and then fusing the sample nonlinear coupling features and the sample core temporal regularity features through a cross-channel fusion layer. The system extracts sample fusion features based on regularity characteristics. Using a gated temporal modeling mechanism, it mines the correlation features between long-term and short-term changes in evaporation within the sample fusion features through multi-layer memory units. These features are then mapped through a fully connected output layer to generate predicted evaporation values ​​for the training phase. The error between the predicted evaporation values ​​and historical actual evaporation values ​​is calculated using a loss function. Based on an error backpropagation mechanism, the system iteratively updates the weights of the multi-layer memory units, the cross-channel fusion weights, and the parameters of the fully connected output layer of the fusion prediction model until the loss function value of the validation set does not decrease for a preset number of consecutive iterations, triggering an early stopping mechanism to halt iteration. Cross-validation is used to verify the performance of the trained fusion prediction model. If the verification accuracy reaches a preset threshold, the training of the fusion prediction model is considered complete.

[0083] Among them, the historical weighted feature sequence fragments are feature sequence fragments generated within a historical period that contain weighted information of core meteorological factors, and are used as meteorological feature inputs for model training; the historical master prediction mode fragments are mode fragments extracted within a historical period that reflect the core temporal features of evaporation, and are used as temporal feature inputs for model training; the training labels are the actual historical evaporation values ​​corresponding to the training data, serving as a benchmark for judging prediction errors during model training; the error backpropagation mechanism is a core training mechanism that calculates the error between the predicted value and the actual value, propagates it backward to each layer of the model, and iteratively updates the parameters to reduce model prediction errors; the early stopping mechanism is a regularization method that stops iterative model training and avoids model overfitting when the loss function value of the validation set does not decrease for a preset number of consecutive rounds; and cross-validation is an evaluation method that comprehensively tests the model's generalization ability and stability by dividing the model into multiple training sets and validation sets for alternating training and evaluation.

[0084] Accordingly, for the embodiments of this disclosure, step 260 may include the following steps: Step 260-1: Using a parallel multi-channel structure, the weighted feature sequence is aligned with the main prediction mode in the time dimension.

[0085] Among them, the parallel multi-channel structure refers to a model input architecture that can simultaneously and independently process different types of feature data (such as meteorological features and time series features). Each feature channel transmits data in parallel and maintains synergy, which can ensure efficient processing and subsequent fusion of multi-source features. Time dimension alignment refers to matching and calibrating the time nodes of the weighted feature sequence and the main prediction mode with a unified timestamp as a benchmark, ensuring that the two types of data correspond one-to-one in the same time dimension and eliminating the problem of time series misalignment.

[0086] In this embodiment of the disclosure, a multi-channel structure capable of parallel processing of multi-source features can be adopted. With a unified time reference as the core, the weighted feature sequence carrying the core meteorological impact information and the main prediction mode reflecting the core temporal pattern of evaporation are synchronously calibrated through adaptation methods such as timestamp matching and data completion. This ensures that the information of the two types of data corresponds to each other at each time node, achieving accurate alignment in the time dimension and laying the foundation for subsequent construction of a unified input tensor and cross-channel feature fusion.

[0087] This technical step, through a parallel multi-channel structure, ensures the high efficiency of multi-source feature processing. At the same time, by aligning the time dimension, it eliminates the temporal misalignment problem between the weighted feature sequence and the main prediction mode, ensuring the temporal consistency of the two types of core input data. This effectively avoids feature fusion deviations caused by data asynchrony. The precisely aligned multi-source data can provide a high-quality input foundation for subsequent tensor construction and cross-channel deep fusion, enabling the fusion prediction model to accurately capture the synergistic influence relationship between meteorological factors and the temporal pattern of evaporation, thereby improving the prediction accuracy and stability of the model.

[0088] Step 260-2: Concatenate the aligned weighted feature sequence with the main prediction mode to construct a unified multi-input tensor.

[0089] Among them, the unified multi-input tensor refers to the tensor data structure that meets the input format requirements of the fusion prediction model and is obtained through splicing operations. It integrates the weighted features of meteorological factors and the core time-series features of evaporation, and can be directly input into the model for calculation.

[0090] In this embodiment of the disclosure, based on the weighted feature sequence and the main prediction mode that have been precisely aligned in the time dimension, the two types of feature data that respectively carry the core meteorological impact information and the core temporal pattern are dimensionally integrated and spliced ​​according to the format requirements of the fusion prediction model for input data. The scattered multi-source features are transformed into tensor forms with unified structure and complete information, and finally a multi-input tensor that can simultaneously provide dual core meteorological and temporal information is constructed, providing standardized input for cross-channel fusion and temporal modeling of the model.

[0091] This technical step, by splicing and aligning the weighted feature sequence with the main prediction mode, can organically integrate the core features of meteorological factors and the core features of evaporation time series, effectively avoiding the fragmentation problem of multi-source input information. The constructed unified multi-input tensor can not only retain the complementary advantages of the two types of features, but also standardize the data format, reduce the processing complexity of multi-source data by the fusion prediction model, and enable the model to efficiently read and integrate key information, laying a solid foundation for subsequent deep fusion of cross-channel features and time series dependency modeling, thereby improving the accuracy and stability of evaporation prediction.

[0092] Step 260-3: Input the multi-input tensor into the trained fusion prediction model, filter the effective dominant factor information in the weighted feature sequence through the model input gate, extract the nonlinear coupling features between dominant factors based on the effective dominant factor information, capture the core temporal regularity features in the main prediction mode through multi-layer memory units, and fuse the nonlinear coupling features and core temporal regularity features through a cross-channel fusion layer to obtain the fusion features.

[0093] The model input gate is one of the core components of the fusion prediction model. It is used to filter out valuable dominant factor information for evaporation prediction from the weighted feature sequence, eliminate redundant interference, and ensure the relevance of the feature input. Effective dominant factor information refers to the non-redundant data information related to the dominant factors that play a key driving role in evaporation changes after being filtered by the input gate in the weighted feature sequence. Nonlinear coupling features refer to the non-linear correlation features formed by the interaction and synergistic influence between dominant factors, which can reflect the complex joint driving effect of meteorological factors on evaporation. Multi-layer memory units are the core components in the fusion prediction model responsible for capturing temporal features (such as the memory units of LSTM). It possesses the ability to mine long-term dependencies in sequences and can accurately extract the core temporal patterns in the main prediction mode. The core temporal pattern features refer to the key temporal characteristics of evaporation contained in the main prediction mode, such as seasonal fluctuations, long-term evolution trends, and periodic changes, which are the core patterns that determine the temporal changes of evaporation. The cross-channel fusion layer is a functional layer in the fusion prediction model used to break the boundary between meteorological feature channels and temporal feature channels, and to achieve deep integration of the two different types of core features. The fusion feature refers to the comprehensive feature that integrates the nonlinear coupling features of the dominant factor and the core temporal pattern features of evaporation after processing by the cross-channel fusion layer, which has both meteorological driving information and temporal evolution patterns.

[0094] In this embodiment of the disclosure, a fusion prediction model trained by multi-input tensors integrating meteorological weighted features and core temporal features can be used. The model accurately selects effective dominant factor information from the weighted feature sequence through an input gate, and further extracts the nonlinear coupling features of the interaction between dominant factors based on this information. At the same time, the model's multi-layer memory unit performs in-depth analysis of the main prediction mode to capture the core temporal regularity features of evaporation contained therein. Finally, the channel barrier between the two types of features is broken through the cross-channel fusion layer, and the extracted nonlinear coupling features and core temporal regularity features are deeply integrated to form a fusion feature that combines meteorological driving effects and temporal evolution patterns.

[0095] This technical step filters effective dominant factor information through model input gates, effectively eliminating redundant interference and improving the effectiveness of feature input. Extracting the nonlinear coupling features of dominant factors accurately characterizes the complex synergistic influence mechanisms among meteorological factors, compensating for the shortcomings of traditional models that neglect factor interactions. Multi-layered memory units capture core temporal regularity features, ensuring an accurate grasp of the essential laws governing the temporal changes in evaporation. The cross-channel fusion layer enables deep complementary integration of two types of core features, generating fused features that combine meteorological driving information with temporal evolution laws. This lays a high-quality feature foundation for subsequent exploration of the correlation between long-term and short-term changes in evaporation, significantly improving the model's adaptability to complex evaporation change mechanisms, thereby ensuring prediction accuracy and stability.

[0096] Step 260-4: Based on the gated time series modeling mechanism built into the fusion prediction model, the long-term and short-term change correlation features of evaporation in the fusion features are mined through multi-layer memory units, and the long-term and short-term change correlation features are mapped to the preliminary evaporation prediction results using the fully connected output layer.

[0097] Among them, the gated temporal modeling mechanism is the core temporal processing mechanism built into the fusion prediction model (such as the LSTM model). Through gating structures such as input gate, forget gate, and output gate, it adaptively controls the transmission and forgetting of information flow to accurately capture the long-term and short-term temporal dependencies. The long-term and short-term change correlation features refer to the inherent correlation between the long-term evolution trend of evaporation (such as seasonal change and annual change) and short-term fluctuations (such as single-day and multi-day small oscillations), which are the key characteristics that determine the temporal changes of evaporation. The fully connected output layer is the core component of the output end of the fusion prediction model. Through linear mapping, it converts the high-dimensional long-term and short-term change correlation features into quantitative results that meet the prediction requirements, realizing the transformation from features to predicted values. The preliminary evaporation prediction result refers to the quantified prediction value of evaporation obtained after mapping by the fully connected output layer without final error calibration or optimization. It is the direct output after the model's temporal modeling and feature mapping, providing the basis for the generation of the subsequent final prediction results.

[0098] In this embodiment of the disclosure, the fusion features of meteorological and temporal core information can be integrated as input. Relying on the gated temporal modeling mechanism built into the fusion prediction model, the intrinsic relationship between the long-term evolution trend and short-term fluctuations of evaporation in the fusion features can be deeply mined through the multi-layer memory unit of the model. Long-term and short-term change correlation features that can fully characterize the temporal dependence relationship are extracted. Then, the high-dimensional long-term and short-term change correlation features are linearly mapped using the fully connected output layer of the model and transformed into preliminary evaporation prediction results that meet the quantification standards, thus completing the core transformation process from feature mining to prediction output.

[0099] This technical step, through the synergistic effect of gated temporal modeling mechanism and multi-layer memory units, can effectively overcome the limitations of traditional models in capturing both long-term and short-term temporal correlations, accurately uncover the inherent laws of long-term and short-term changes in evaporation, and ensure the depth and accuracy of temporal modeling. The fully connected output layer can achieve efficient and accurate mapping from high-dimensional correlation features to quantitative prediction results, which can not only retain the key information of core features, but also ensure the quantitative rationality of prediction results, providing a high-quality foundation for the optimization of the final evaporation prediction results, and significantly improving the model's adaptability to complex temporal changes and the reliability of prediction results.

[0100] Step 260-5: Based on the error distribution pattern and prediction deviation characteristics recorded during the training of the fusion prediction model, the preliminary evaporation prediction results are calibrated to obtain the final evaporation prediction results.

[0101] Among them, the error distribution law refers to the statistical distribution characteristics (such as normal distribution, skewed distribution, etc.) of the error between the predicted value and the actual value recorded during the training process of the fusion prediction model, reflecting the overall trend and distribution pattern of the model prediction error; the prediction deviation characteristics of different seasons refer to the specific patterns of prediction deviation caused by seasonal differences when the model makes predictions in different seasons (such as the deviation characteristics caused by high evaporation in summer, and the deviation performance corresponding to low evaporation in winter, etc.), reflecting the influence of seasonal factors on prediction error; the final evaporation prediction result refers to the accurate prediction result after calibration, with the error further reduced and more in line with the actual evaporation change pattern, which is the final output of the model prediction process.

[0102] In this embodiment of the disclosure, the preliminary evaporation prediction results output by the prediction model can be integrated as a basis. Based on the error distribution patterns (such as the statistical distribution pattern and magnitude range of errors) recorded during the model training phase, and combined with the specific performance characteristics of prediction deviations under different seasons, the preliminary prediction results can be calibrated point by point or in batches through targeted error correction algorithms (such as compensation formulas based on deviation characteristics, error distribution adaptation and adjustment, etc.). This corrects the systematic errors and seasonally related deviations, and finally obtains evaporation prediction results with higher accuracy and more in line with the actual situation.

[0103] This technical step utilizes the error distribution patterns and seasonal deviation characteristics accumulated during model training to accurately calibrate the preliminary prediction results. This effectively corrects potential systematic errors and seasonal biases in the model, compensating for the accuracy limitations caused by relying solely on model feature mapping. The calibrated final prediction results are more consistent with the actual evaporation variation patterns, especially adapting to the differences in evaporation characteristics in different seasons. This significantly improves the accuracy and reliability of the prediction results, providing more accurate decision-making basis for engineering applications such as water resource allocation and irrigation management.

[0104] In summary, the technical solution in this application, by constructing a multivariate regression conversion model through screening meteorological factors that are significantly correlated with heterogeneous evaporation data, can achieve accurate correction and integration of observation data from heterogeneous evaporation pans, effectively eliminating systematic biases in data from different types of evaporation pans. The constructed continuous standardized target evaporation sequence can provide a high-quality data foundation for subsequent analysis. By screening the correlation of target meteorological factors, calculating weights based on a support vector machine model using radial basis function kernel functions, and screening dominant factors to generate a weighted feature sequence, redundant interference from weakly correlated factors can be effectively eliminated, reducing model training noise and significantly improving the effectiveness of feature inputs and the model's generalization ability. Furthermore, by using a variational mode decomposition algorithm to process the target evaporation sequence... Through row decomposition and multidimensional evaluation, the extracted master prediction mode can accurately capture the core patterns of evaporation time series, effectively identifying multi-frequency components, seasonal characteristics, and abrupt change points, thus solving the problem that traditional models struggle to handle non-stationary sequences. Finally, through a fusion prediction model containing multi-layer memory units, cross-channel fusion layers, and fully connected output layers, parallel multi-channel alignment of weighted features and master prediction modes, cross-channel depth and time series dependency modeling can be achieved. This fully integrates the nonlinear coupling effects of meteorological factors with the time series evolution of evaporation. Combined with early stopping mechanisms and cross-validation during model training, model performance can be guaranteed, ultimately significantly improving the accuracy and stability of evaporation prediction and meeting the reliability requirements of practical engineering applications.

[0105] Furthermore, as Figure 1 and Figure 2 The specific implementation of the method shown in this embodiment provides an evaporation prediction device, such as... Figure 3 As shown, the device includes: a construction module 31, a filtering module 32, an extraction module 33, and an input module 34.

[0106] Module 31 can be used to convert and correct observation data from heterogeneous evaporating dishes to construct a continuous and standardized target evaporation sequence. The screening module 32 can be used to identify the influence weight of the target meteorological factors on evaporation based on the correlation between the target evaporation sequence and meteorological factors, and to screen out the dominant factors from the target meteorological factors based on the influence weight and generate a weighted feature sequence. Extraction module 33 can be used to perform mode decomposition on the target evaporation sequence and extract the main prediction mode that reflects the core time series characteristics; Input module 34 can be used to train a fusion prediction model by inputting weighted feature sequences and main prediction modality, and output the evaporation prediction result after parallel multi-channel data alignment and tensor construction, deep fusion of cross-channel features and gated temporal dependency modeling.

[0107] In some embodiments of this application, the construction module 31 can be specifically used to screen meteorological factors that are significantly related to the differences in heterogeneous evaporating pan observation data as adjustment variables, construct a multivariate regression conversion model containing adjustment variables, perform conversion correction on the heterogeneous evaporating pan observation data based on the multivariate regression conversion model, integrate the converted and corrected heterogeneous evaporating pan observation data, and obtain a continuous standardized target evaporation sequence.

[0108] In some embodiments of this application, the screening module 32 can be specifically used to perform correlation analysis between the target evaporation sequence and multiple preset meteorological factors, and to screen multiple target meteorological factors whose correlation is greater than a preset correlation threshold. The preset meteorological factors include at least air temperature, relative humidity, wind speed, sunshine duration, air pressure, surface temperature, and solar radiation. Using the target evaporation sequence as the output variable and the target meteorological factors as the input variables, a support vector machine regression model based on the radial basis function kernel function is established. The kernel parameters and penalty factors of the support vector machine regression model are optimized, and the optimized support vector machine model is used to train and cross-validate each target meteorological factor. Using the support vector machine regression model after training and cross-validation, the influence weight of each target meteorological factor on evaporation is calculated.

[0109] In some embodiments of this application, the screening module 32 can be further used to normalize the influence weight of each target meteorological factor to obtain a standardized weight; to determine the target meteorological factor whose standardized weight is greater than a preset weight screening threshold as the dominant factor; to extract the original observation data sequence corresponding to each dominant factor, and to perform a dot product operation between the original observation data sequence and the corresponding standardized weight to obtain the weighted data sequence of each dominant factor; and to concatenate the weighted data sequences of all dominant factors along the time dimension to form a unified weighted feature sequence.

[0110] In some embodiments of this application, the extraction module 33 can be specifically used to decompose the target evaporation sequence using a variational mode decomposition algorithm to obtain multiple stationary intrinsic mode functions (IMF) components, each IMF component corresponding to different frequency characteristics of the target evaporation sequence; perform multidimensional evaluation on each IMF component, including calculating the energy proportion of each IMF component, verifying stationarity using the ADF test method, and analyzing the correlation with the target evaporation sequence using the Pearson correlation coefficient method; based on the multidimensional evaluation results, select target IMF components whose energy proportion is higher than a preset energy threshold, whose stationarity test results meet a preset stationarity standard, and whose correlation with the target evaporation sequence is higher than a preset correlation coefficient threshold; and determine the target IMF components as the main prediction modes that can reflect the core temporal characteristics of evaporation.

[0111] In some embodiments of this application, the fusion prediction model includes a multi-layer memory unit, a cross-channel fusion layer, and a fully connected output layer. The input module 34 can specifically be used to align the weighted feature sequence with the main prediction mode in a time dimension using a parallel multi-channel structure; concatenate the aligned weighted feature sequence with the main prediction mode to construct a unified multi-input tensor; input the multi-input tensor into the trained fusion prediction model, filter effective dominant factor information from the weighted feature sequence through the model input gate, extract nonlinear coupling features between dominant factors based on the effective dominant factor information, capture core temporal regularity features in the main prediction mode through the multi-layer memory unit, and fuse the nonlinear coupling features and core temporal regularity features through the cross-channel fusion layer to obtain fused features; based on the gated temporal modeling mechanism built into the fusion prediction model, mine the long-term and short-term variation correlation features of evaporation in the fused features through the multi-layer memory unit, and map the long-term and short-term variation correlation features to preliminary evaporation prediction results using the fully connected output layer; calibrate the preliminary evaporation prediction results based on the error distribution pattern recorded during the training of the fusion prediction model and the prediction deviation features of different seasons to obtain the final evaporation prediction results.

[0112] In some embodiments of this application, such as Figure 4 As shown, the device also includes: a training module 35; Training module 35 can be used to select historical weighted feature sequence fragments and corresponding historical main prediction mode fragments as training data, and simultaneously use the actual historical evaporation values ​​corresponding to each main prediction mode fragment as training labels. The training data and training labels are divided into training set and validation set according to a preset ratio. The training set data and corresponding training labels are input into the constructed fusion prediction model. The model input gate extracts the sample nonlinear coupling features between the dominant factors based on the historical weighted feature sequence. The multi-layer memory unit captures the core temporal regularity features of the samples in the historical main prediction mode. Then, the cross-channel fusion layer fuses the sample nonlinear coupling features and the sample core temporal regularity features to obtain the sample fusion. Features: Based on a gated temporal modeling mechanism, the long-term and short-term variation correlation features of evaporation in the sample fusion features are mined through multi-layer memory units. The evaporation prediction values ​​during the training phase are generated by mapping through a fully connected output layer. The error between the predicted evaporation values ​​and the actual historical evaporation values ​​is calculated by combining the loss function. Based on the error backpropagation mechanism, the weights of the multi-layer memory units, the cross-channel fusion weights, and the parameters of the fully connected output layer of the fusion prediction model are iteratively updated until the loss function value of the validation set does not decrease for a preset number of consecutive rounds, triggering an early stopping mechanism to stop the iteration. The performance of the trained fusion prediction model is verified by cross-validation. If the verification accuracy reaches a preset threshold, the fusion prediction model is judged to have completed training.

[0113] It should be noted that other corresponding descriptions of the functional units involved in the evaporation prediction device provided in this embodiment can be found in [reference needed]. Figure 1 and Figure 2 The corresponding descriptions in [the document] will not be repeated here.

[0114] Based on the above, Figure 1 and Figure 2 Accordingly, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method. Figure 1 and Figure 2 The method for predicting evaporation is shown.

[0115] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause an electronic device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0116] Based on the above, Figure 1 and Figure 2 The method shown, and Figure 3 , 4To achieve the above objectives, the present application also provides an electronic device, specifically a personal computer, tablet computer, server, or other network device, as shown in the virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 and Figure 2 The method for predicting evaporation is shown.

[0117] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0118] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0119] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0121] This invention constructs a multivariate regression conversion model by screening meteorological factors that are significantly correlated with heterogeneous evaporation data. This model enables accurate correction and integration of observational data from heterogeneous evaporation pans, effectively eliminating systematic biases in data from different types of evaporation pans. The constructed continuous standardized target evaporation sequence provides a high-quality data foundation for subsequent analysis. By screening the correlation of target meteorological factors, calculating weights using a support vector machine model based on radial basis function kernel functions, and selecting dominant factors to generate a weighted feature sequence, redundant interference from weakly correlated factors can be effectively eliminated, reducing model training noise and significantly improving the effectiveness of feature inputs and the model's generalization ability. The target evaporation sequence is decomposed using a variational mode decomposition algorithm. With multidimensional evaluation, the extracted master prediction mode can accurately capture the core laws of evaporation time series, effectively identify multi-frequency components, seasonal characteristics and abrupt change points, and solve the problem that traditional models cannot handle non-stationary sequences. Finally, through the fusion prediction model containing multi-layer memory units, cross-channel fusion layers and fully connected output layers, parallel multi-channel alignment of weighted features and master prediction modes, cross-channel depth and time series dependency modeling can be achieved, fully integrating the nonlinear coupling influence of meteorological factors and the time series evolution law of evaporation. Combined with the early stopping mechanism and cross-validation in the model training process, the model performance can be guaranteed, and the accuracy and stability of evaporation prediction can be significantly improved, meeting the requirements of engineering practical applications for prediction reliability.

[0122] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0123] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for predicting evaporation, characterized in that, include: The observation data of heterogeneous evaporating dishes were converted and corrected to construct a continuous and standardized target evaporation sequence. Based on the correlation between the target evaporation sequence and meteorological factors, the influence weight of the target meteorological factors on evaporation is identified, and the dominant factor is screened from the target meteorological factors based on the influence weight and a weighted feature sequence is generated. The target evaporation sequence is subjected to mode decomposition to extract the main prediction mode that reflects the core temporal characteristics; The fusion prediction model, trained by the weighted feature sequence and the main prediction modality input, outputs the evaporation prediction result after parallel multi-channel data alignment and tensor construction, cross-channel feature deep fusion, and gated temporal dependency modeling.

2. The method according to claim 1, characterized in that, The conversion and correction of observation data from heterogeneous evaporating dishes to construct a continuous and standardized target evaporation sequence includes: Meteorological factors that are significantly correlated with the differences in observation data of heterogeneous evaporation pans are selected as adjustment variables, and a multiple regression conversion model including the adjustment variables is constructed. Based on the aforementioned multivariate regression conversion model, the observation data of the heterogeneous evaporating dish are converted and corrected. The converted and corrected observation data of the heterogeneous evaporating dish are then integrated to obtain a continuous and standardized target evaporation sequence.

3. The method according to claim 1, characterized in that, Based on the correlation between the target evaporation sequence and meteorological factors, the influence weights of the target meteorological factors on evaporation are identified, including: The target evaporation sequence is subjected to correlation analysis with multiple preset meteorological factors, and multiple target meteorological factors with a correlation greater than a preset correlation threshold are selected. The preset meteorological factors include at least air temperature, relative humidity, wind speed, sunshine duration, air pressure, surface temperature and solar radiation. Using the target evaporation sequence as the output variable and the target meteorological factors as the input variables, a support vector machine regression model based on radial basis kernel function is established; The kernel parameters and penalty factor of the support vector machine regression model are optimized, and the optimized support vector machine model is used to train and cross-validate each of the target meteorological factors. The influence weight of each target meteorological factor on evaporation is calculated using a support vector machine regression model that has been trained and cross-validated.

4. The method according to claim 3, characterized in that, Based on the influence weights, dominant factors are selected from the target meteorological factors, and a weighted feature sequence is generated, including: The influence weights of each of the target meteorological factors are normalized to obtain standardized weights; The target meteorological factors whose standardized weights are greater than the preset weight screening threshold are identified as the dominant factors; Extract the original observation data sequence corresponding to each of the dominant factors, and perform a dot product operation between the original observation data sequence and the corresponding standardized weight to obtain the weighted data sequence of each of the dominant factors; The weighted data sequences of all the dominant factors are concatenated along the time dimension to form a unified weighted feature sequence.

5. The method according to claim 1, characterized in that, The target evaporation sequence is subjected to mode decomposition to extract the main prediction mode that reflects the core time-series characteristics, including: The target evaporation sequence is decomposed using a variational mode decomposition algorithm to obtain multiple stationary intrinsic mode functions (IMF) components, each of which corresponds to a different frequency feature of the target evaporation sequence. A multidimensional evaluation is performed on each of the intrinsic mode functions (IMF) components, including calculating the energy proportion of each IMF component, verifying stationarity using the ADF test, and analyzing the correlation with the target evaporation sequence using the Pearson correlation coefficient method. Based on the multidimensional evaluation results, target intrinsic mode function (IMF) components with an energy ratio higher than a preset energy threshold, a stationarity test result that meets a preset stationarity standard, and a correlation with the target evaporation sequence higher than a preset correlation coefficient threshold are selected. The target intrinsic mode function (IMF) component is determined as the master prediction mode that can reflect the core temporal characteristics of evaporation.

6. The method according to claim 1, characterized in that, The fusion prediction model includes a multi-layer memory unit, a cross-channel fusion layer, and a fully connected output layer. The fusion prediction model, trained by the weighted feature sequence and the main prediction modality input, is then subjected to parallel multi-channel data alignment and tensor construction, cross-channel feature deep fusion, and gated temporal dependency modeling to output evaporation prediction results, including: A parallel multi-channel structure is adopted to align the weighted feature sequence with the main prediction mode in the time dimension; The aligned weighted feature sequence is concatenated with the main prediction mode to construct a unified multi-input tensor. The fusion prediction model trained by the multi-input tensor input is used to filter the effective dominant factor information in the weighted feature sequence through the model input gate. Based on the effective dominant factor information, the nonlinear coupling features between the dominant factors are extracted. At the same time, the core temporal regularity features in the main prediction mode are captured through the multi-layer memory unit. The nonlinear coupling features and the core temporal regularity features are fused through the cross-channel fusion layer to obtain the fused features. Based on the gated time series modeling mechanism built into the fusion prediction model, the long-term and short-term change correlation features of evaporation in the fusion features are mined through the multi-layer memory unit, and the long-term and short-term change correlation features are mapped to the preliminary evaporation prediction results using the fully connected output layer. Based on the error distribution pattern and prediction deviation characteristics recorded during the training of the fusion prediction model, the preliminary evaporation prediction results are calibrated to obtain the final evaporation prediction results.

7. The method according to claim 6, characterized in that, The method also includes a training method for the fusion prediction model: Historical weighted feature sequence segments and corresponding historical main prediction mode segments are selected as training data. Simultaneously, the actual historical evaporation values ​​corresponding to each main prediction mode segment are used as training labels. The training data and training labels are divided into training set and validation set according to a preset ratio. The training set data and corresponding training labels are input into the completed fusion prediction model. The nonlinear coupling features between the dominant factors are extracted through the model input gate based on the historical weighted feature sequence. The core temporal regularity features of the samples in the historical main prediction mode are captured through the multi-layer memory unit. Then, the nonlinear coupling features and the core temporal regularity features of the samples are fused through the cross-channel fusion layer to obtain the sample fusion features. Based on the gated time series modeling mechanism, the long-term and short-term variation correlation features of evaporation in the sample fusion features are mined through multi-layer memory units. The evaporation prediction value in the training stage is generated by mapping through a fully connected output layer. The error between the evaporation prediction value and the actual historical evaporation value is calculated by combining the loss function. Based on the error backpropagation mechanism, the weights of the multi-layer memory units, the cross-channel fusion weights, and the parameters of the fully connected output layer of the fusion prediction model are iteratively updated until the loss function value of the validation set does not decrease for a preset number of consecutive rounds, triggering the early stop mechanism to stop the iteration. The performance of the trained fusion prediction model is verified by cross-validation. If the verification accuracy reaches a preset threshold, the training of the fusion prediction model is considered complete.

8. An evaporation prediction device, characterized in that, include: The module is used to convert and correct the observation data of heterogeneous evaporating dishes and construct a continuous and standardized target evaporation sequence. The screening module is used to identify the influence weight of the target meteorological factors on evaporation based on the correlation between the target evaporation sequence and meteorological factors, and to screen out the dominant factors from the target meteorological factors based on the influence weight and generate a weighted feature sequence. The extraction module is used to perform mode decomposition on the target evaporation sequence and extract the main prediction mode that reflects the core time series characteristics; The input module is used to input the weighted feature sequence and the fusion prediction model trained by the main prediction modality into the fusion prediction model, and output the evaporation prediction result after parallel multi-channel data alignment and tensor construction, cross-channel feature deep fusion and gated temporal dependency modeling.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images