Data driving-based electricity-to-water coefficient composite model and prediction method
By constructing a data-driven composite model and comprehensive evaluation system, and combining multi-dimensional evaluation indicators and genetic algorithm optimization, the limitations of a single model and the defects of the evaluation system in the prediction of the electricity-to-water ratio are solved, achieving higher accuracy and reliability in prediction results, which are applicable to the estimation of the electricity-to-water ratio in different regions and scenarios.
Patent Information
- Application Number
- CN202511462523.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-14
- Publication Date
- 2026-01-13
AI Technical Summary
Existing technologies for predicting water efficiency using electricity conversion factors suffer from limitations such as single-model limitations, a lack of scientific evaluation systems, and insufficient quantification of the uncertainty of prediction results. These issues result in insufficient prediction accuracy and generalization ability, making it impossible to adapt to changes in data characteristics across different regions and scenarios.
A data-driven composite model is adopted, combining EPC-STIN neural network, CNN neural network, RBF model and regional prediction model. By optimizing the index weights through multi-dimensional evaluation indicators and genetic algorithm, a comprehensive evaluation system is constructed. The interpolation range and Gaussian weight allocation are dynamically adjusted. Combined with multi-dimensional evaluation indicators and prediction interval coverage, the multi-dimensional characteristics of the electro-water conversion coefficient are modeled and predicted.
It significantly improves the accuracy and generalization ability of prediction based on the electro-water conversion factor, adapts to changes in data characteristics under different scenarios, provides a more comprehensive evaluation basis, helps users scientifically quantify the uncertainty of prediction results, and enhances the credibility of decision-making.
Smart Images

Figure CN121327602A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electric water conversion factor prediction technology, specifically involving a data-driven composite model and prediction method for electric water conversion factor. Background Technology
[0002] Factors such as pump performance, well age, pump age, groundwater level depth, well location and spatial distribution, well-pump matching, well completion quality, and hydrogeological conditions all significantly impact well output and the hydroelectric conversion coefficient. Effective resource data analysis and organization, along with the establishment of a data observation and hydroelectric conversion coefficient calculation model, are crucial for supporting and promoting the practical application of hydropower big data integration and for carrying out agricultural water resource tax reform.
[0003] Current technologies focus on using a single machine learning model, such as a backpropagation (BP) neural network. BP neural networks attempt to capture the non-linear relationship between electricity consumption and water consumption by learning from large amounts of historical data. They utilize the connection weights between neurons to simulate the mapping between input and output, adjusting the weights through multiple iterations of training to achieve better predictive results. However, this method primarily focuses on the numerical features and patterns of the data, neglecting the underlying geospatial distribution characteristics. In practical applications, the electricity-to-water conversion factor is often influenced by spatial factors such as geographical location and topography. BP neural networks cannot effectively process this spatial information, thus performing poorly in regionalized predictions.
[0004] Some existing technologies rely on traditional spatial interpolation methods, such as inverse distance weighted interpolation (IDW) and kriging interpolation. IDW, based on the principle of distance decay, assumes that sample points closer to the point to be predicted have a greater influence, and assigns weights for interpolation by calculating the distance between the sample point and the point to be predicted. Kriging interpolation, on the other hand, is based on regionalized variable theory, using a variogram to describe the correlation of spatial variables, thereby achieving optimal unbiased estimation. However, these methods primarily focus on geographic location and do not adequately consider the impact of engineering parameters such as well depth, power, and water level on the electricity-to-water conversion factor. Furthermore, they typically employ fixed parameters and model structures, making it difficult to adapt to changes in data characteristics across different regions and scenarios.
[0005] In addition, traditional statistical models, such as linear regression models, are also used in existing technologies. Linear regression models assume a linear relationship between variables and use the least squares method to fit the data to obtain a regression equation for prediction. However, in reality, the relationship between the electricity conversion coefficient and various influencing factors is often non-linear, and linear regression models cannot accurately capture this complex relationship, resulting in limited prediction accuracy.
[0006] In summary, the existing technology has the following drawbacks:
[0007] Limitations of a single model: A single model can only process data from a specific perspective and cannot comprehensively consider various influencing factors. For example, while BP neural networks excel at handling nonlinear relationships, they lack the ability to process spatial information; and while spatial interpolation methods can handle geospatial data, they are not sensitive enough to engineering parameters. This limitation makes the model have poor generalization ability when facing complex and ever-changing real-world situations, and the prediction results are prone to bias.
[0008] The evaluation system suffers from serious flaws: most existing technologies employ only a single evaluation metric, such as root mean square error (RMSE), to measure model performance. This single-metric evaluation method is too one-sided, only reflecting the model's performance in a certain aspect and failing to comprehensively assess the model's quality. For example, RMSE only reflects the average error between predicted and actual values, but it cannot reflect the stability and reliability of the prediction results. Moreover, existing evaluation systems often set metric weights based on subjective experience, lacking scientific rigor and objectivity, leading to inaccurate model selection and potentially choosing models that perform well in some aspects but have poor overall performance.
[0009] Traditional spatial interpolation methods lack flexibility and adaptability: these methods typically employ fixed parameters and model structures, failing to dynamically adjust to data characteristics across different regions and scenarios. For example, the distance weight powers in the IDW method are usually fixed; in areas with uneven sample distribution or complex terrain, this fixed weight allocation may lead to inaccurate interpolation results. Kriging interpolation requires pre-assuming a variogram model, but in reality, the correlations between spatial variables can be very complex, and the pre-assumed model may fail to accurately describe these correlations, thus affecting interpolation accuracy.
[0010] Insufficient quantification of forecast uncertainty: Most existing technologies only provide point forecasts, i.e., a single predicted value, without quantifying the uncertainty of the forecast. In practical applications, the uncertainty of the forecast is crucial information, helping decision-makers better assess risk. The lack of calculation and analysis of forecast intervals in existing technologies prevents decision-makers from understanding the reliability of the forecast, increasing the risk of their decisions. Summary of the Invention
[0011] To address the aforementioned problems in the existing technology, this invention provides a data-driven composite model and prediction method based on the electro-water conversion coefficient. The technical problem to be solved by this invention is achieved through the following technical solution:
[0012] A data-driven composite model and prediction method based on the electro-water conversion coefficient includes:
[0013] S100 collects multi-source heterogeneous data related to the water pump and preprocesses it to obtain preprocessed data;
[0014] S200: Train multiple different prediction models using the training set data in the preprocessed data to obtain multiple trained prediction models, and test the multiple trained prediction models using the test set data in the preprocessed data to obtain the predicted value of the electricity-water conversion coefficient output by each prediction model; wherein, the multiple different prediction models include at least the EPC-STIN model and the regionalized prediction model.
[0015] S300: For each prediction model, calculate the model's metrics and standardized composite score based on the predicted and actual values output by the model.
[0016] S400 selects the prediction model with the highest standardized comprehensive score from all trained prediction models as the optimal prediction model, and uses this optimal prediction model to predict new multi-source heterogeneous data to obtain the water conversion coefficient.
[0017] Beneficial effects:
[0018] 1. This invention compares and analyzes spatial interpolation, neural network, and regionalized prediction models, and performs differentiated modeling based on the multi-dimensional characteristics of the electro-water conversion coefficient (spatial correlation, nonlinear electro-water correlation, and engineering parameters). It overcomes the limitations of single models, adapts to different data characteristics, significantly improves prediction accuracy and generalization ability, and is applicable to electro-water conversion coefficient estimation in various scenarios.
[0019] 2. The EPC-STIN neural network proposed in this invention innovatively constructs an architecture of "engineering physical constraints - cross-dimensional feature interaction": it extracts spatial-temporal features through 3D convolution and dynamic LSTM, and designs "space-engineering coupling gate triggered by well depth > 100m" and "temporal-engineering coupling gate triggered by activation time ≥ 5 years", embedding engineering logic such as "well depth × head → coefficient reduction" and "equipment aging → increased power fluctuation" into the network structure; at the same time, through physical constraint loss function (coefficient range 0.5~8) and dynamic adaptation mechanism, it solves the problems of "weak generalization ability and poor physical consistency" of traditional hybrid models, and provides a dedicated deep network solution for the three-dimensional coupling prediction of "space-engineering-temporal" based on the electric-to-water coefficient.
[0020] 3. This invention designs an interpolation method (regional prediction model) based on spherical distance and dynamic radius optimization. By adaptively adjusting the interpolation range and Gaussian weight allocation, it strengthens the influence of local samples. Compared with traditional spatial interpolation with fixed parameters, this invention can dynamically adjust the strategy according to the data distribution, significantly improving the accuracy and adaptability of spatial interpolation, and is especially suitable for areas with uneven sample distribution or complex terrain.
[0021] 4. This invention proposes a comprehensive evaluation method that combines multi-dimensional evaluation indicators (RMSPE, RMSE, PICP, etc.). It uses a genetic algorithm (GA) to dynamically optimize indicator weights, achieving objective model selection. This solves the problem of arbitrary subjective weighting in traditional evaluations, ensuring that the comprehensive score reflects the model's balanced performance in terms of error accuracy, prediction interval reliability, and other aspects, thus improving the scientific rigor of model selection.
[0022] 5. This invention introduces the Predicted Interval Coverage (PICP) and Average Width (PINAW) metrics, combined with traditional error metrics (RMSPE, RMSE, etc.), to evaluate the model from both accuracy and reliability dimensions. This provides a more comprehensive evaluation basis for predicting the electro-water conversion factor, helping users scientifically quantify the uncertainty of prediction results and improve the credibility of decision-making.
[0023] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0024] Figure 1 This invention provides a flowchart illustrating a data-driven composite model and prediction method based on the electro-water conversion coefficient.
[0025] Figure 2 This is a schematic diagram showing the comparison between the predicted values and the actual values of the optimal regional prediction model provided by the present invention on the test set.
[0026] Figure 3 The diagram shows the application effect of the optimal model regional prediction model under the dataset provided by this invention in geospatial space.
[0027] Figure 4 This is a scatter plot of the predicted and actual values of the EPC-STIN model in this embodiment of the invention.
[0028] Figure 5 This is a scatter plot of the predicted and actual values of the CNN model in this embodiment of the invention.
[0029] Figure 6 This is a residual distribution diagram of the EPC-STIN model in this embodiment of the invention;
[0030] Figure 7 This is a residual distribution diagram of the CNN model in this embodiment of the invention;
[0031] Figure 8 This is a two-dimensional interpolation plot of the Kriging model;
[0032] Figure 9 This is a two-dimensional interpolation plot of the RBF model;
[0033] Figure 10This is a frequency distribution chart of the electro-water conversion coefficient in the dataset provided by this invention;
[0034] Figure 11 This is a schematic diagram illustrating the relationship between longitude, latitude, and hydroelectric conversion factor in the dataset provided by this invention.
[0035] Figure 12 The dataset provided by this invention is a scatter plot of water level and coefficients.
[0036] Figure 13 The dataset provided by this invention is a scatter plot of well depth and coefficients;
[0037] Figure 14 The dataset provided by this invention is a schematic diagram showing the distribution of electro-hydraulic coefficients under different head conditions;
[0038] Figure 15 This is a heatmap showing the correlation between the residual variables and the water conversion coefficient in the dataset provided by this invention. Detailed Implementation
[0039] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0040] To address the problems of existing technologies, this invention proposes a data-driven composite model and prediction method for the electro-hydraulic coefficient. It establishes a composite model for the electro-hydraulic coefficient encompassing three main categories: neural network models, spatial interpolation models, and regionalized prediction models (among which the EPC-STIN model and the regionalized prediction model are innovative architectures for electro-hydraulic prediction models). Multiple evaluation indicators, including error metrics (RMSE, MAE, MSE), relative error metrics (RMSPE), and interval prediction metrics (PICP, PINAW), are used to construct a comprehensive evaluation system to evaluate the prediction results. This allows for the output of the most accurate prediction model for different datasets. The optimal prediction model is then used to predict the electro-hydraulic coefficient of ordinary wells.
[0041] Combination Figure 1 This invention provides a data-driven composite model and prediction method based on the electro-water conversion coefficient, including:
[0042] S100 collects multi-source heterogeneous data related to the water pump and preprocesses it to obtain preprocessed data; among which, spatial information includes longitude and latitude, engineering parameters include well depth, water level, pipe diameter, power, head and water pump start-up time, and target values include the electro-water conversion factor.
[0043] S200: Train multiple different prediction models using the training set data in the preprocessed data to obtain multiple trained prediction models, and test the multiple trained prediction models using the test set data in the preprocessed data to obtain the predicted value of the electricity-water conversion coefficient output by each prediction model; wherein, the multiple different prediction models include at least the EPC-STIN model and the regionalized prediction model.
[0044] S300: For each prediction model, calculate the model's metrics and standardized composite score based on the predicted and actual values output by the model.
[0045] S400 selects the prediction model with the highest standardized comprehensive score from all trained prediction models as the optimal prediction model, and uses this optimal prediction model to predict new multi-source heterogeneous data to obtain the water conversion coefficient.
[0046] The composite model effectively combines data processing and prediction. In practical applications, the data is constantly updated and changing. New data, including spatial information, engineering parameters, and actual measurements of the electro-water conversion factor, are collected periodically. The prediction model is retrained using the updated data, and its parameters are adjusted to adapt to the data changes. Finally, the optimal prediction model is used to predict the electro-water conversion factor for ordinary wells. This dynamic process provides a scientific basis for predicting the electro-water conversion factor in agricultural water management. Each step is closely integrated with data processing and model application to ensure the reliability and accuracy of the prediction results.
[0047] In one specific embodiment of the present invention, S100 includes:
[0048] S110 collects multi-source heterogeneous data containing spatial information, engineering parameters, and target values;
[0049] S120, perform missing value processing and outlier processing on the multi-source heterogeneous data in sequence to obtain the processed data;
[0050] The missing value handling process is as follows: Missing fields in the data are identified through statistical analysis or visualization. For fields with low missing value rates, records containing missing values are directly deleted. For fields with high missing value rates, imputation is performed using the mean, median, interpolation methods (such as time series interpolation and spatial interpolation), or prediction based on machine learning models (such as KNN and regression models). The outlier handling process is as follows: Outliers are detected using statistical methods and machine learning algorithms. A small number of outliers are directly deleted. Outliers caused by data collection errors are corrected through interpolation or model fitting.
[0051] S130: Select data from the processed data whose correlation with the target value exceeds a predetermined correlation threshold, and then standardize and normalize them to obtain preprocessed data.
[0052] This step uses domain knowledge and relevance analysis to select features highly correlated with the target value. Then, data standardization and normalization are performed. Standardization (Z-score standardization) and normalization (Min-Max scaling) eliminate dimensional differences, improving model training efficiency and stability. The standardization and normalization formulas are as follows:
[0053] ;
[0054] ;
[0055] in, The mean, Standard deviation The maximum value in the data. It is the minimum value.
[0056] In one specific embodiment of the present invention, S200 includes:
[0057] S210, the preprocessed data is divided into training set and test set according to a predetermined ratio;
[0058] This step can divide the preprocessed data into training and testing sets, usually randomly in a ratio of 7:3 or 8:2.
[0059] S220 involves training multiple different prediction models using a training set, obtaining multiple trained prediction models, and then testing these models using a test set to obtain the predicted value of the electro-water conversion coefficient output by each model. These multiple prediction models include EPC-STIN neural networks, CNN neural networks, RBF models, Kriging interpolation models, and regionalization prediction models.
[0060] This application imports pre-processed data in real time. This data is then used to train and test the composite model, with dynamic adjustments to the model parameters during training. The composite model comprises multiple prediction models, primarily including neural network models, spatial interpolation models, and regionalized prediction models. Among these, the EPC-STIN neural network model and the regionalized prediction model are innovative electro-hydraulic prediction models. Finally, a comprehensive index is used to evaluate the trained model and select the optimal model. This optimal model is then used to predict the coefficients of ordinary wells, i.e., the electro-hydraulic coefficients.
[0061] (1) EPC-STIN Neural Network: A composite deep neural network that integrates engineering physical constraints and spatial-temporal features to solve the problems of traditional models ignoring physical laws and poor generalization ability of multiple models in parallel. The core principle is to integrate the physical influence mechanism of the electro-water conversion coefficient (spatial proximity, correlation of engineering parameters, temporal stability) into the deep neural network through "structural embedding + constraint optimization" to achieve dual-driven prediction of "data-driven + physical guidance".
[0062] The EPC-STIN neural network consists of an engineering-physical linkage preprocessing layer, a three-dimensional factor interaction extraction layer, a physical constraint cross-dimensional fusion layer, an engineering-adaptive prediction layer, and a loss function.
[0063] 1) Engineering physical linkage pretreatment layer: The missing well depth and water level are supplemented by linear regression.
[0064] 2) Three-dimensional factor interaction extraction layer:
[0065] Spatial Feature Branch: Extract the three-dimensional spatial pattern of latitude, longitude and well depth through 3D convolution, and embed the physical law that "the greater the well depth, the higher the spatial weight".
[0066] Engineering Features Branch: Engineering Features: By enhancing the influence of key parameters such as head and power through the attention mechanism, it conforms to the measured law that "head is negatively correlated with coefficient".
[0067] Timing characteristics: The long-term trend of water pump aging is captured by dynamic window LSTM, and the engineering logic of "the longer the equipment is used, the wider the timing window" is embedded.
[0068] 3) Physical constraint cross-dimensional fusion layer:
[0069] Space-Engineering Coupling Gating: Triggering condition h>100m, coupling characteristics are as follows ,in For element-wise product, For well depth dimension features, This refers to the head characteristics.
[0070] Time-engineering coupling gating: Trigger condition activation time t>5 years, coupling characteristic is ,in, The power fluctuation timing characteristics, This represents the power characteristics.
[0071] Final fusion features:
[0072] 4) Project adaptability prediction: based on the predicted value of the electro-hydraulic conversion factor. , ,in , These are the output layer parameters.
[0073] 5) Loss function:
[0074] Data loss: ;
[0075] Interval constraints: ;
[0076] Physical constraints: ;
[0077] Weight: , .
[0078] (2) CNN Neural Network: In the prediction field, one-dimensional convolutional neural networks (1D-CNN) are considered the mainstream method for constructing feature extraction models. This network architecture identifies the inherent patterns and correlations of time-series data through local feature extraction operations, integrates nonlinear activation functions and downsampling modules to improve the model's representation ability and perform feature space compression, thereby achieving hierarchical extraction of abstract features. Its mathematical model is expressed as follows:
[0079]
[0080] In the formula: f is the input; f is the activation function; Used to perform invalid convolutions without zero padding; It is the first The nucleus of the i-th neuron in the layer; It is the first The output of the i-th neuron in the layer; It is the first Bias of k neurons in a layer
[0081] (3) RBF Model: The radial basis function model approximates the target function through a linear combination of a set of radial basis functions. Commonly used radial basis functions include Gaussian functions and multiple quadratic functions. This model considers latitude and longitude and well depth as inputs during training, constructing a three-dimensional prediction model using latitude, longitude, and well depth as three dimensions. Verification shows that considering these three features improves accuracy compared to considering only latitude and longitude. Predicted values The calculation formula is as follows:
[0082]
[0083] in It is a radial basis function. These are sample points. It's the weight.
[0084] (4) Kriging interpolation model: This is a geostatistical interpolation method based on regionalized variable theory. It determines the weights by calculating the spatial covariance (variogram) between sample points, thereby performing optimal unbiased estimation. The predicted value is... The calculation formula is as follows:
[0085]
[0086] Among the sample points The point to be measured is Weight The Kriging equations are obtained by solving the system of variograms. Model construction. This model takes latitude and longitude, well depth, water level, and head as inputs during model training, and a spherical model as the variation function.
[0087] (7) Regionalized prediction model: In response to the technical pain points of the electric water conversion coefficient, namely "geographical distribution dominated by spherical characteristics, strong correlation with engineering parameters (well depth), and uneven spatial distribution of samples", this invention proposes an adaptive interpolation method based on spherical distance correction, multi-objective dynamic radius optimization and coupling with engineering parameters. The specific innovations are as follows:
[0088] Based on the distribution of sample points around the point to be predicted (center) (within the circle), the interpolation radius is dynamically adjusted. The predicted value is obtained by weighting the sample points within the radius using distance. The calculation formula is as follows:
[0089]
[0090] Among them, weight , It is the distance between the point to be measured and the sample point. This is the weighting function (Gaussian weighting function), and the formula for the Gaussian weighting function is as follows:
[0091]
[0092] This model considers latitude and longitude during training, finding the optimal radius that minimizes the weighted mean squared error (WMSE) by traversing a series of possible radius values. The predicted value is then obtained by taking a distance-weighted average of the sample points within the radius.
[0093] This implementation proposes an interpolation method based on spherical distance and dynamic radius optimization. By adaptively adjusting the interpolation range and Gaussian weight allocation, it enhances the influence of local samples. Compared to traditional spatial interpolation with fixed parameters, this model can dynamically adjust the strategy according to the data distribution, significantly improving the accuracy and adaptability of spatial interpolation, and is particularly suitable for areas with uneven sample distribution or complex terrain.
[0094] This implementation method includes spatial interpolation models (such as RBF models, Kriging interpolation, etc.), neural network models (such as EPC-STIN, CNN), and regionalized prediction models, which perform differentiated modeling for the multi-dimensional characteristics of the electro-water conversion coefficient (spatial correlation, nonlinear electro-water correlation, engineering parameters). It overcomes the limitations of a single model, adapts to different data characteristics, significantly improves prediction accuracy and generalization ability, and is suitable for estimating the electro-water conversion coefficient in different scenarios.
[0095] In one specific embodiment of the present invention, S300 includes:
[0096] S310, for each prediction model, calculate multiple evaluation metrics for the prediction model using the predicted and actual values output by the prediction model; among which, the multiple evaluation metrics include error-related evaluation metrics and prediction interval metrics. Error-related evaluation metrics include root mean square percentage error, root mean square error, mean absolute error, and mean relative error; prediction interval metrics include prediction interval coverage and prediction interval average width.
[0097] Error-related indicators:
[0098] RMSPE (Root Mean Square Percentage Error): Measures the relative error between the predicted and actual values, reflecting the accuracy of the prediction. The formula is:
[0099]
[0100] in It is the actual value. It is a predicted value. It refers to the number of samples.
[0101] RMSE (Root Mean Square Error): Measures the average error between the predicted and actual values. The formula is:
[0102]
[0103] MAE (Mean Absolute Error): Calculates the average of the absolute errors between predicted and actual values. The formula is:
[0104]
[0105] MRE (Mean Relative Error): Reflects the average deviation of predicted values from actual values. The formula is:
[0106]
[0107] Forecast interval indicators:
[0108] PICP (Predicted Interval Coverage): Measures the proportion of predicted intervals that contain the true values, reflecting the reliability of the prediction. The formula is:
[0109]
[0110] in, and They are the first The lower and upper bounds of the prediction interval for each sample. It is an indicator function.
[0111] PINAW (Average Width of Forecast Intervals): Calculates the average width of all forecast intervals, reflecting the breadth or narrowness of the forecast intervals. The formula is:
[0112]
[0113] This implementation introduces the Predicted Interval Coverage (PICP) and Average Width (PINAW) metrics, combined with traditional error metrics (RMSPE, RMSE, etc.), to evaluate the model from both accuracy and reliability dimensions. This provides a more comprehensive evaluation basis for predicting the electro-water conversion factor, helping users scientifically quantify the uncertainty of prediction results and improve the credibility of decision-making.
[0114] Existing technologies often employ single indicators or subjective weighting in predictive model evaluation, leading to biased model selection. This invention, however, constructs a comprehensive evaluation system that introduces multi-dimensional indicators (RMSPE, RMSE, PICP, etc.) to balance error accuracy and the reasonableness of the prediction range. It eliminates the influence of dimensions through indicator positiveization and standardization, and utilizes a genetic algorithm (GA) to dynamically optimize indicator weights, avoiding subjective assignment bias. This system ensures the objectivity and scientific rigor of model evaluation, resulting in optimal models with greater practical application value.
[0115] S320 calculates a standardized composite score for each prediction model, based on the weights of each evaluation metric and multiple evaluation metrics.
[0116] In one specific embodiment of the present invention, S320 includes:
[0117] S321, Standardize and positively process each evaluation indicator to obtain the positively processed evaluation indicators;
[0118] S322, determine the weight of each evaluation indicator;
[0119] This step uses the sum of the comprehensive scores of all prediction models as the fitness function; it employs a genetic algorithm (GA) with the fitness function as the optimization objective, and iteratively searches for an optimal weight vector that maximizes the standardized comprehensive score of all prediction models through genetic operations such as selection, crossover, and mutation; this optimal weight is then used as the weight for each evaluation index. The standardized comprehensive score is the weighted sum of the standardized values of each index.
[0120] S323, For each prediction model, and based on the weights of each evaluation indicator and the positively processed multiple evaluation indicators, calculate the standardized composite score of the prediction model. The standardized composite score is expressed by the formula:
[0121]
[0122] In the formula, This represents the weight of each evaluation indicator j. This represents the standardized evaluation index, where m represents the number of evaluation indexes.
[0123] This implementation proposes a comprehensive evaluation method that combines multi-dimensional evaluation indicators (RMSPE, RMSE, PICP, etc.). It uses a genetic algorithm (GA) to dynamically optimize indicator weights, achieving objective model selection. This solves the problem of arbitrary subjective weighting in traditional evaluations, ensuring that the comprehensive score reflects the model's balanced performance in terms of error accuracy, prediction interval reliability, and other aspects, thus improving the scientific rigor of model selection.
[0124] The comprehensive evaluation system described above allows for a comprehensive and objective assessment of the performance of each model, enabling the selection of the optimal model based on the overall score. The evaluation indicators measure the model's predictive accuracy and reliability from different perspectives. Standardization and positive conversion of the indicators eliminate the influence of dimensions and direction, while the weights determined by the genetic algorithm ensure a reasonable weight allocation for each indicator in the comprehensive evaluation.
[0125] Instance data verification
[0126] Data from actual measurements at a specific location was used. After data preprocessing, the model's inputs included longitude, latitude, water level, pipe diameter, power, head, and pump start-up time. The output was the electro-hydraulic conversion factor. The model was built using Python, and the measured electro-hydraulic conversion data was divided into training and testing sets in an 8:2 ratio. The training set was used to train the model, and the testing set was used to test the model and obtain corresponding predicted values. Relevant indicators were calculated based on the actual and predicted values. Table 1 shows example data from the dataset, and Table 2 shows the optimal weights determined by the GA (Graphical Analysis) algorithm.
[0127] Table 1
[0128]
[0129] Table 2
[0130]
[0131] Genetic Algorithm (GA) simulates the processes of selection, crossover, and mutation in biological evolution to find a set of weights that maximize the overall score of all models. This is the optimal weight. Table 2 shows that MRE (0.425) has the highest weight, indicating that the mean relative error plays a central role in the overall evaluation; the relative deviation between the model's predicted values and the actual values is a key factor in distinguishing model quality. RMSE (0.336) has the second highest weight, reflecting the model's high attention to the overall level of prediction error; models with larger root mean square errors will be significantly penalized. Other indicators have relatively low weights, indicating their weaker influence on the overall evaluation.
[0132] Table 3
[0133]
[0134] As shown in Table 3, the regionalized prediction model is the optimal model. Its accuracy indicators, such as RMSPE, RMSE, MAE, and MRE, are the lowest among all models, demonstrating the best prediction accuracy and stability. PICP=0.929 is close to 0.95, and PINAW=1.083, which is the narrowest among models with the same coverage level. Its overall score of 0.992 is the highest among all models, showing a balanced performance in both accuracy and reliability. The EPC-STIN model is the second best model. Although its accuracy indicators are slightly higher than the regionalized prediction model, they are significantly better than other models. In terms of reliability indicators, PICP is consistent with the regionalized prediction model, and PINAW=1.267 remains competitive. Its overall score of 0.967 reflects its effective fitting ability to the spatial-engineering-temporal coupling relationship of the electro-water conversion coefficient.
[0135] The CNN model significantly outperforms Kriging and RBF models in accuracy metrics, demonstrating the advantage of neural networks in fitting complex nonlinear relationships. Only its RMSPE is slightly higher than the Kriging model, making it more suitable for nonlinear scenarios involving engineering parameters. The Kriging model has a high PICP but a large PINAW and high uncertainty, making it more suitable for scenarios dominated by pure spatial correlation. The RBF model performs the worst across all metrics, presumably due to insufficient matching between its function and the "geographic-engineering" coupling distribution characteristics represented by the electricity-water conversion coefficient. In summary, regionalized prediction models are the optimal choice, EPC-STIN has alternative value in complex scenarios, and CNN, Kriging, and RBF require targeted optimization before being selected for specific scenarios.
[0136] refer to Figures 2 to 3 As shown, Figure 2This diagram illustrates the comparison between the predicted and actual values of the optimal regional prediction model provided by this invention on the test set. Figure 3 The diagram shows the application effect of the optimal model regional prediction model under the dataset provided by this invention in geospatial space. Figure 2 The blue dots represent the correspondence between the predicted and true values for each sample, while the red dashed line represents the diagonal line where, ideally, the predicted and true values are exactly equal. Analysis shows that most blue dots are closely distributed near the red dashed line, indicating high prediction accuracy. Furthermore, the distribution trend of the dots is largely consistent with the ideal diagonal line, suggesting that the model maintains stable prediction performance across different true value ranges without systematic bias. In addition, no obvious outliers were observed in the scatter plot, indicating that overfitting or underfitting did not occur during model training, demonstrating its ability to accurately capture the inherent patterns in the data and maintain prediction stability even with complex data distributions. Figure 3 In the diagram, different colored dots represent specific attribute values of the sample points, such as the electrochemical conversion factor, while the green circles indicate the dynamic search radius considered during interpolation. This radius is adjusted according to the distribution and density of the sample points to ensure the accuracy and spatial continuity of the interpolation. The longitude range of the sample points in the figure is 109.30 to 110.0, and the latitude range is 34.0 to 35.0. The interpolation results are more accurate in densely populated areas of sample points, while some uncertainty is shown in sparse areas.
[0137] refer to Figures 4 to 7 As shown, where, Figure 4 This is a scatter plot of the predicted and actual values of the EPC-STIN model in this embodiment of the invention. Figure 5 These are scatter plots showing the predicted and true values of the CNN model in this embodiment of the invention. In these two plots, the horizontal axis represents the true value, the vertical axis represents the predicted value, and the red dashed line represents the reference line (y=x) where the predicted value and the true value are perfectly consistent in an ideal situation. By comparison, it can be found that the predicted value of the EPC-STIN model is closer to the reference line than the true value, indicating that its prediction accuracy is higher. In contrast, the scatter plot distribution of the CNN model is relatively scattered, especially in the region of high true value, where the predicted value deviates from the true value to a greater extent, showing that the prediction performance of the CNN model is not as good as that of the EPC-STIN model when processing high-value data.
[0138] Figure 6 This is a residual distribution diagram of the EPC-STIN model in this embodiment of the invention. Figure 7 This is a residual distribution diagram of the CNN model in this embodiment of the invention. The residual refers to the difference between the predicted value and the true value; the distribution of the residuals reflects the distribution characteristics of the model's prediction error. Figure 6In the EPC-STIN model, the residual distribution is relatively concentrated, with most residual values close to 0 and a relatively symmetrical distribution, indicating that the model's prediction error is small and stable. In contrast, Figure 7 The residual distribution of the CNN model is relatively dispersed and has a certain degree of skewness, especially in the region where the residual value is positive, the frequency of the residual is relatively high. This indicates that the CNN model may systematically overestimate the true value in some cases.
[0139] refer to Figures 8 to 9 As shown, Figure 8 This is a two-dimensional interpolation plot of the Kriging model. Figure 9 This is a two-dimensional interpolation plot of the RBF model. Figure 8 The Kriging model in the model shows good local fitting ability and can capture local changes around the data points well, but it may experience interpolation instability in areas with few data points. Figure 9 The RBF model interpolation plots show relatively complex spatial variability and can reflect the nonlinear relationships of the data well, but overfitting may occur in regions with very dense data points. Overall, the Kriging model is slightly better than the RBF model.
[0140] refer to Figure 10 , Figure 10 The histogram shows the frequency distribution of the dataset "electricity-to-water conversion coefficient". The histogram reveals a unimodal distribution, with data clustering around a certain coefficient value to form a distinct peak. The right tail of the distribution is relatively long, while the left side is steep, exhibiting a right-skewed (positively skewed) shape. This means that most of the data is concentrated in the lower coefficient value range, while higher coefficient values occur less frequently.
[0141] refer to Figure 11 , Figure 11 This is a scatter plot showing the relationship between the longitude, latitude, and the electrochemical coefficient of the dataset. The colors change from purple to yellow, representing different coefficient values. Yellow indicates a higher coefficient value (close to 7), and purple indicates a lower coefficient value (close to 1). The high and low value regions show a certain degree of clustering, indicating that the coefficients are spatially correlated.
[0142] refer to Figure 12 and Figure 13 , Figure 12 This is a scatter plot of the dataset's water level and coefficients. Figure 13 This is a scatter plot of the dataset's well depth versus coefficients. The coefficients show a clear decreasing trend as the water level increases. Figure 12This indicates that the coefficient values are higher when the water level is low, and lower when the water level is high. Data points are more densely distributed in the low water level region (0-40), and the coefficient values range widely, from 1 to 8. In the high water level region (above 80), data points are sparser, and the coefficient values are relatively low, mostly concentrated below 1.
[0143] Figure 13 This indicates a certain negative correlation between well depth and the coefficient. As well depth increases, the coefficient tends to decrease gradually, but this relationship is not as pronounced as the relationship between water level and the coefficient. In shallower well areas (0-40), the coefficient values are more dispersed, ranging from approximately 1 to 7. As well depth increases, the coefficient values gradually decrease, and after well depth exceeds 100, the coefficient values mostly concentrate around 1.
[0144] refer to Figure 14 , Figure 14 This is a schematic diagram showing the distribution of electro-hydraulic coefficients under different head conditions in the dataset. From Figure 14 The data shows that the median of the electro-water conversion coefficient fluctuates with increasing head. The median is relatively high in the low and medium head regions, but decreases in the high head region. This indicates that at low and medium head levels, the water resource extraction per unit of electricity consumption is relatively low, resulting in a higher electro-water conversion coefficient; while at high head levels, the water resource extraction per unit of electricity consumption is relatively high, resulting in a lower electro-water conversion coefficient. The dispersion of the electro-water conversion coefficient data varies significantly across different head levels. At some head points (e.g., 36.0, 73.01), the data box is longer, indicating a wider data distribution and greater differences between samples; while at other head points (e.g., 150.0, 270.0), the data box is shorter, indicating more concentrated data.
[0145] refer to Figure 15 , Figure 15This is a heatmap showing the correlation between the remaining variables in the dataset and the coefficient of water converted from electricity. This plot illustrates the strength of the correlations between multiple variables. Pipe diameter shows low correlations with other variables (power, pump start-up time, and value), with the highest correlation (0.17) with the value (coefficient of water converted from electricity), but this is still a weak positive correlation, indicating that pipe diameter has a relatively small impact on other variables. The correlation between power and pump start-up time is close to zero, indicating almost no linear correlation between them. The correlation between power and the value (coefficient of water converted from electricity) is -0.45, a moderate negative correlation, indicating that higher power may correlate with a smaller value. The correlation between pump start-up time and the value (coefficient of water converted from electricity) is -0.089, a weak negative correlation, indicating that pump start-up time has a small impact on the value, and the trend is that the longer the pump start-up time, the slightly smaller the coefficient may be. Correlation between the value and other variables: The correlation between the value and pipe diameter, power, and pump start-up time is 0.17, -0.45, and -0.089, respectively, indicating that the value is mainly negatively correlated with power, while the correlation with pipe diameter and pump start-up time is weak.
[0146] This invention addresses the problem of predicting the electricity-to-water conversion factor in agricultural water management. It innovatively constructs the EPC-STIN model and a regionalized prediction model, and establishes a composite prediction system incorporating multiple models. A comprehensive evaluation system is used to fully assess the performance of each model. The constructed comprehensive evaluation system overcomes the limitations of single-indicator evaluation by using multi-index quantitative analysis (including error-related and interval prediction indicators) and genetic algorithm (GA) to optimize weights, providing a scientific basis for model selection. The composite model system demonstrates strong applicability in predicting the electricity-to-water conversion factor in agricultural water management. It can select the optimal model for accurate prediction based on different data characteristics. It is not only applicable to electricity-to-water conversion factor prediction but can also be extended to the evaluation of other geospatial data-driven agricultural models. This system promotes the transformation of water resource management from "experience-driven" to "data-driven," providing strong methodological support for precision irrigation decision-making.
[0147] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.
[0148] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A data-driven composite model and prediction method based on the electro-water conversion coefficient, characterized in that, include: S100 collects multi-source heterogeneous data related to the water pump and preprocesses it to obtain preprocessed data; S200, using the training set data in the preprocessed data, train multiple different prediction models to obtain multiple trained prediction models, and use the test set data in the preprocessed data to test the multiple trained prediction models to obtain the predicted value of the electro-water conversion coefficient output by each prediction model; wherein, the multiple different prediction models include at least the EPC-STIN model and the regionalized prediction model. S300: For each prediction model, based on the predicted and actual values output by the prediction model, calculate the indicators of the prediction model, determine the weight of each indicator according to the comprehensive evaluation system, and calculate the comprehensive score. S400 selects the prediction model with the highest comprehensive score from all trained prediction models as the optimal prediction model, and uses the optimal prediction model to predict new multi-source heterogeneous data to obtain the water conversion coefficient.
2. The data-driven composite model and prediction method based on the electro-water conversion coefficient as described in claim 1, characterized in that, S100 includes: S110 collects multi-source heterogeneous data containing spatial information, engineering parameters, and target values; S120, the multi-source heterogeneous data is sequentially processed for missing values and outliers to obtain the processed data; S130, Data with a correlation to the target value exceeding a predetermined correlation threshold is selected from the processed data, and then standardized and normalized to obtain preprocessed data.
3. The data-driven composite model and prediction method based on the electro-water conversion coefficient as described in claim 2, characterized in that, The spatial information includes longitude and latitude, the engineering parameters include well depth, water level, pipe diameter, power, head, and pump start-up time, and the target value includes the electro-water conversion factor.
4. The data-driven composite model and prediction method based on the electro-water conversion coefficient as described in claim 1, characterized in that, S200 includes: S210, the preprocessed data is divided into a training set and a test set according to a predetermined ratio; S220, multiple different prediction models are trained using the training set to obtain multiple trained prediction models, and the multiple trained prediction models are tested using the test set to obtain the predicted value of the electro-water conversion coefficient output by each prediction model.
5. The data-driven composite model and prediction method based on the electro-water conversion coefficient according to claim 1 or 4, characterized in that, The regionalized prediction model is a dynamic radius optimization interpolation model based on spherical distance. It determines the optimal radius by minimizing the weighted mean square error by traversing the radius values, and calculates the predicted value using a Gaussian weight function that integrates well depth / head parameters. The EPC-STIN model is an engineering physical constraint-spatial-temporal interactive composite deep network. It extracts the three-dimensional spatial features of latitude, longitude and well depth through 3D convolutional layers, uses a dynamic window LSTM to capture the temporal features of pump start-up time and power fluctuations, and designs physical constraint gating units for well depth-head coupling gate and aging-power coupling gate. It embeds engineering physical laws such as the electro-water conversion coefficient approaching stability with increasing well depth and the influence coefficient of equipment aging power fluctuations into the network structure, realizing cross-dimensional interactive fusion and end-to-end prediction of spatial, engineering and temporal features.
6. The data-driven composite model and prediction method based on the electro-water conversion coefficient according to claim 1, characterized in that, The S300 includes: S310, For each prediction model, use the predicted values and actual values output by the prediction model to calculate multiple evaluation metrics for the prediction model. S320, For each prediction model, and based on the weights of each evaluation indicator and the multiple evaluation indicators, calculate the standardized comprehensive score of the prediction model.
7. The data-driven composite model and prediction method based on the electro-water conversion coefficient according to claim 6, characterized in that, The multiple evaluation indicators include error-related evaluation indicators and prediction interval indicators. The error-related evaluation indicators include root mean square percentage error, root mean square error, mean absolute error, and mean relative error. The prediction interval indicators include prediction interval coverage and prediction interval average width.
8. The data-driven composite model and prediction method based on the electro-water conversion coefficient according to claim 6, characterized in that, S320 includes: S321, Standardize and positively process each evaluation indicator to obtain the positively processed evaluation indicators; S322, determine the weight of each evaluation indicator; S323. For each prediction model, and based on the weights of each evaluation indicator and multiple evaluation indicators after positive processing, calculate the standardized composite score of the prediction model.
9. The data-driven composite model and prediction method based on the electro-water conversion coefficient according to claim 8, characterized in that, The determination of the weights for each evaluation indicator includes: The fitness function is the sum of the combined scores of all prediction models. Using a genetic algorithm with fitness function maximization as the optimization objective, the optimal weight vector is continuously searched through selection, crossover and mutation operations to maximize the standardized comprehensive score of all prediction models. The optimal weights are used as the weights for each evaluation index.
10. The data-driven composite model and prediction method based on the electro-water conversion coefficient according to claim 6, characterized in that, The standardized composite score is expressed by the formula: ; In the formula, Indicates the first The weight of each evaluation indicator, This represents the standardized evaluation indicators. This indicates the number of evaluation indicators.
Citation Information
Cited By
Coastal erosion risk multi-dimensional evaluation method based on machine learning
CN121765480A
