A multi-source fusion method for predicting groundwater levels

Abstract

This invention, entitled "A Multi-Source Fusion Groundwater Level Prediction Method," belongs to the field of groundwater level prediction technology. The technical problem it aims to solve is that existing technologies suffer from a lack of core geological parameters in their data sources, insufficient optimization and failure to capture time dependence in their feature systems, and crude data preprocessing mechanisms. The key technical solution is as follows: First, the prediction time node is determined based on multi-source monitoring data. Then, a prediction feature set containing basic and lagging features is constructed. Missing value imputation and Min-Max normalization preprocessing are performed on this set. Two independently trained random forest regression models are used to generate preliminary predictions of the highest and lowest water levels. Finally, after logical verification and correction, the prediction result, including key borehole information and preprocessed records, is output.

Description

Technical Field

[0001] This invention belongs to the field of groundwater level prediction technology, specifically relating to a multi-source fusion groundwater level prediction method. Background Technology

[0002] Under the dual pressures of global climate change and intensified human activities, identifying groundwater conditions and monitoring groundwater level changes are crucial aspects of engineering construction, groundwater resource management, and disaster prevention. However, the movement of groundwater is complex and influenced by numerous factors. Given the current conditions of intensified global climate change and frequent human engineering activities, accurately and efficiently predicting groundwater levels based on geological and environmental conditions has become a major technical challenge.

[0003] Groundwater level prediction technology has gradually evolved from traditional physical models and statistical analysis methods towards multi-source data fusion and artificial intelligence-driven approaches. Traditional physical models rely on detailed hydrogeological parameters, limiting their applicability in data-scarce areas. While single statistical or machine learning models are easy to operate, they suffer from incomplete data source coverage, coarse data preprocessing, incomplete feature systems, and imprecise initial prediction processes, making them unsuitable for complex groundwater systems involving multiple coupled factors such as geology, meteorology, and hydrology. Currently, multi-source data fusion has become a core direction for improving prediction accuracy; however, existing technologies mostly remain at the level of surface data overlay, failing to achieve deep integration throughout the entire process. Furthermore, they lack systematic solutions for data quality optimization, refined feature processing, and standardized initial predictions, thus failing to meet the high-precision prediction needs under complex geological conditions.

[0004] Relevant patent documents retrieved:

[0005] (1) Country of publication: China, Publication number: CN120746015A. This document discloses an AI intelligent groundwater level dynamic prediction system based on multi-source data fusion. It collects multi-source data such as meteorological data, hydrological data, regional groundwater dynamic monitoring data, and building life cycle structural health monitoring data. It determines the factor weights through multi-criteria decision analysis and multi-agent decision-making mechanism, and uses deep learning algorithms to predict groundwater levels.

[0006] (2) Country of publication: China, Publication number: CN202310523817.6. This document discloses a groundwater level prediction method. Based on the hydrological and hydrogeological conditions of the study area, a conceptual model and a mathematical model are constructed. A deep neural network is trained by analytical solution, and then the parameters are fine-tuned to establish a migration model for prediction. It depends on the fixed hydrogeological conditions of the study area.

[0007] Relevant non-patent literature retrieved: (1) Journal name: Groundwater Dynamics, document name: Theory and application of groundwater dynamics, volume number: third edition, publication date: 2010. This document discloses a physical process driven model, which describes groundwater movement based on the mass conservation theory and partial differential equations, and relies on detailed hydrogeological parameters.

[0008] (2) Journal Name: Document Title: Volume No.: 577, Publication Date: 2019. This paper discloses a groundwater level prediction method based on machine learning algorithms such as ANN and SVM, which relies only on historical data features and does not integrate core geological parameters.

[0009] (3) Journal Name: The paper, titled "Ground water level prediction / forecasting and assessment of uncertainty using sgsandarima models: a case study in the Baurua quifer system (Brazil)," published in 2018, discloses the application of the ARIMA model, which is based solely on historical water level time series predictions and does not consider time-dependent features.

[0010] (4) Book title: Distributed watershed hydrological model, document title: Principles and applications of distributed hydrological models, publication date: 2004. This document discloses distributed hydrological models such as VIC, which require massive amounts of high-resolution topographic, soil and meteorological data, making them difficult to apply in areas with scarce data.

[0011] The prior art represented by the aforementioned documents has at least the following unresolved technical problems or defects: (1) Incomplete data source coverage and lack of core geological parameters: Traditional methods often rely on single-type data. Comparison documents such as CN120746015A and non-patent literature such as the Journal of Hydrology do not cover key parameters that directly affect rock permeability and rock mass integrity, such as rock quality index (RQD) and Lurong value. It is difficult to reflect the multi-factor coupling effect of geology, meteorology, and hydrology. The relevant evidence is that the existing technology does not incorporate them into the multi-source data system and cannot fully capture the impact of geological conditions on water level.

[0012] (2) The data preprocessing mechanism is crude, and the data quality is difficult to guarantee: Existing technologies only use simple filling or deletion to handle missing values ​​and outliers, without designing hierarchical processing logic for random missing values ​​and continuous missing values ​​in time series, and without establishing a standardized format unification and rationality verification process. Relevant evidence is that improper handling of missing values ​​and outliers in existing technologies will lead to a sharp drop in the model's generalization ability, continuous missing values ​​in time series will destroy the continuity of data, and inconsistent formats will directly lead to the interruption of the prediction process.

[0013] (3) Incomplete feature system and insufficient capture of time dependence: Existing technologies do not construct lagged features to capture the impact of previous data on the current water level, and do not perform refined feature optimization, resulting in differences in feature magnitude interfering with model weight allocation and redundant information affecting model learning efficiency. Relevant evidence is that groundwater level changes have significant time dependence, and existing technologies do not design lagged features; at the same time, there is a lack of systematic feature optimization mechanisms, which cannot guarantee the effectiveness and adaptability of the feature set.

[0014] (4) The initial prediction process is not refined, and its operability and accuracy are insufficient: Existing technologies do not clearly define key steps such as the rules for determining the time node of the initial prediction, the method for constructing prediction feature matching, and the standard for format uniformity and adaptation. The comparison document CN120746015A does not disclose the specific implementation logic of the initial prediction. Relevant evidence is that the initial prediction needs to ensure that the features are completely consistent with the model training format. Existing technologies lack a systematic design for predictive feature construction, missing feature filling, and normalization of numerical features, which leads to deviations in prediction results and makes it difficult to directly apply them.

[0015] In view of this, the present invention is hereby proposed. Summary of the Invention

[0016] To address the aforementioned technical problems in existing technologies, this invention provides a multi-source fusion groundwater level prediction method, which solves the problems of existing technologies such as lack of core geological parameters in data sources, failure to capture time dependence in feature systems and insufficient optimization, and crude data preprocessing mechanisms.

[0017] To achieve the above objectives, the technical solution of the present invention is as follows: A multi-source fusion method for predicting groundwater levels includes: S1. Determine the prediction time node: Based on the multi-source groundwater monitoring data, extract the latest monitoring month as the base month, and calculate the next month after the base month as the prediction time node; S2. Construct a predictive feature set: For each monitored borehole, extract the dynamic monitoring data and borehole static parameters of the baseline month, and construct a predictive feature set that includes basic features and lag features. S3. Preprocessing the predicted feature set: Fill in the missing values ​​in the predicted feature set and normalize the numerical features to ensure that the feature set is compatible with the input format of the training model. S4. Generate preliminary prediction values ​​for dual objectives: Call two independently trained regression models, input the preprocessed prediction feature sets respectively, and generate preliminary prediction values ​​for the highest water level and the lowest water level corresponding to the prediction time node. S5. Verification, Correction and Output: Perform logical verification on the preliminary prediction values ​​of the dual targets, correct prediction values ​​that do not conform to logic, and output the preliminary prediction results that include key information of the monitoring borehole and the corrected prediction values.

[0018] Furthermore, the specific rules for calculating the predicted time nodes are as follows: If the base month is December, the prediction time point is the base year plus 1, and the month is fixed at 1; If the base month is from January to November, the prediction time node is the month with the base year unchanged and the month is the base month plus 1. The prediction time node is uniformly converted into a string in the format "YYYY-MM".

[0019] Furthermore, the static parameters of the borehole mentioned in step S2 include: unique borehole identifier, latitude, longitude, elevation, rock quality index, and Lloyd's value; The dynamic monitoring data for the baseline month includes rainfall, highest water level, and lowest water level.

[0020] Furthermore, the basic features include: the monthly value corresponding to the predicted time point, the borehole static parameters, and the rainfall in the baseline month; The lagging characteristics include rainfall, highest water level, and lowest water level for the base month.

[0021] Furthermore, the specific method for filling in missing values ​​in step S3 is as follows: If there are missing feature values ​​in the predicted feature set, the missing feature values ​​include: random missing values ​​or missing first values ​​of time series. The average value of the same type of feature values ​​of all other monitored boreholes in the same area is used to fill the missing values. If there are no other monitored boreholes in the area, the historical statistical mean of this feature is used to fill the feature, and the feature value is marked as a fill identifier in the feature set.

[0022] Furthermore, the specific content of the filling identifier includes: the name of the missing feature, the area range used for filling, the number of boreholes used in the calculation of the mean, and the filling time.

[0023] Furthermore, in step S3, the normalization process employs the Min-Max normalization method, specifically mapping the numerical features to the [0,1] interval. The mapping formula is as follows:

[0024] in, These are the original eigenvalues. This is the minimum value of the feature. This is the maximum value of the feature, eliminating the impact of differences in the magnitude of different features on the model weight allocation.

[0025] Furthermore, in step S4, both independently trained regression models are random forest regression models, with the model parameters set as follows: ; One model is trained on the historical highest water level to output a preliminary prediction of the highest water level; the other model is trained on the historical lowest water level to output a preliminary prediction of the lowest water level.

[0026] Furthermore, the specific process of logical verification in step S5 is as follows: Determine whether the preliminary forecast value of the lowest water level is greater than the preliminary forecast value of the highest water level. If so, directly swap the preliminary forecast values ​​of the lowest and highest water levels, and after correction, satisfy the condition that the preliminary forecast value of the lowest water level is less than or equal to the preliminary forecast value of the highest water level.

[0027] Furthermore, the key information monitored in step S5 includes: the unique identifier of the borehole, the latitude coordinates of the borehole, the longitude coordinates of the borehole, the elevation of the borehole, and the predicted time node. The preliminary prediction results also include preprocessing records of the predicted feature set, which include: missing value imputation, normalized feature names and processing parameters.

[0028] The beneficial effects of this invention are as follows: (1) Dual-target independent prediction improves targeting and accuracy. Two independently trained prediction models are used to predict the highest and lowest water levels respectively, adapting to the differences in the influencing factors of the two types of water level indicators. This avoids the accuracy loss caused by a single model predicting two targets at the same time, making the prediction results more consistent with the changing characteristics of different water level indicators. (2) Multi-source feature integration enhances the comprehensiveness of prediction. It integrates multi-source information such as borehole static parameters, meteorological data, and historical water levels to construct a complete feature set containing basic features and lag features. It covers the multi-factor coupling influence of geology, meteorology, and hydrology, and effectively captures the time dependence of water level changes, solving the problem that traditional models ignore external driving factors or time correlation patterns. (3) The preprocessing mechanism improves data adaptability and anti-interference ability. By filling missing values ​​with the mean of similar features in the region and standardizing and normalizing numerical features, the method can still run stably in actual scenarios where data is missing or has different formats. This significantly reduces the stringent requirements on the quality of the original data and improves the environmental adaptability of the method. (4) Logical verification ensures the rationality of the prediction results. The numerical logical verification and automatic correction steps for the lowest and highest water levels are added to avoid outputting contradictory prediction results and ensure that the prediction results conform to the basic laws of hydrology, thereby improving the reliability and usability of the prediction results. Attached Figure Description

[0029] Figure 1 A flowchart of a multi-source fusion groundwater level prediction method provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0031] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0032] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0033] Example See Figure 1 , Figure 1 This is a flowchart of a multi-source fusion groundwater level prediction method proposed in this invention. Specific steps may include: S1. Determine the prediction time node: Based on the multi-source groundwater monitoring data, extract the latest monitoring month as the base month, and calculate the next month after the base month as the prediction time node; extract the latest monitoring month from the multi-source groundwater monitoring data ( Using the base month as the reference month, the predicted time points are calculated according to the following rules ( ): If the base month is December, the prediction time node is the base year plus 1, and the month is fixed at 1 (e.g., base month 2026-12 → prediction time node 2027-01). If the base month is from January to November, the predicted time node is the month in which the base year remains unchanged and the month is the month value of the base month plus 1 (e.g., base month 2026-05 → predicted time node 2026-06). The predicted time nodes are uniformly converted into strings in the format "YYYY-MM".

[0034] S2. Constructing a predictive feature set: For each monitored borehole, extract the dynamic monitoring data and borehole static parameters of the baseline month to construct a predictive feature set containing basic features and lag features; specifically including: S21. According to the unique identifier of the borehole ( The process iterates through all monitoring boreholes within the region. For each borehole, the latest monitoring record corresponding to the baseline month is selected, and the dynamic monitoring data and borehole static parameters are extracted from that record. The dynamic monitoring data includes monthly rainfall, maximum water level, and other parameters. ) and minimum water level ( Static parameters include the borehole unique identifier (…). Latitude, longitude, elevation, rock quality index (RQD), and Lürson value (LQD) ).

[0035] S22. Create a prediction feature set for each monitoring borehole corresponding to the prediction time node. This feature set includes basic features and lag features: Basic feature: The month value corresponding to the predicted time point ( The values ​​range from 1 to 12), borehole static parameters, and rainfall for the baseline month. If the rainfall for the baseline month is missing, the historical average for the same period of the borehole will be used to fill the gap. Lag characteristics: Rainfall (rainfall_lag1), highest water level (water_level_max_lag1), and lowest water level (water_level_min_lag1) for the base month; if the base month is the first monitoring record for this borehole (without previous data), the lag characteristic values ​​are filled with the average of the lag characteristics of all other monitoring boreholes in the same area, and a filling indicator is marked.

[0036] S3. Preprocessing the predicted feature set: Missing values ​​in the predicted feature set are imputed, and numerical features are normalized to ensure the feature set matches the input format of the training model; specifically including: S31. Missing Value Handling: If there are missing feature values ​​in the predicted feature set (e.g., a borehole has no RQD data), the missing feature values ​​include: Random missing values ​​(such as no RQD data for a certain borehole, missing rainfall data for the baseline month, etc.): are filled by the mean of the same type of characteristic values ​​from all other monitored boreholes in the same area; For time series with missing first values ​​(e.g., the base month is the first monitoring record for this borehole, and there is no lagged feature data): fill with the mean of the lagged features of all other monitoring boreholes in the same area during the same period; S32. Normalization Processing: Perform Min-Max normalization processing on the numerical features in the predicted feature set, mapping the feature values ​​to the [0,1] interval. The mapping formula is as follows:

[0037] in, These are the original eigenvalues. This is the minimum value of the feature. The maximum value of this feature is used to eliminate the influence of differences in the magnitude of different features on the model weight allocation, ensuring that the feature set is compatible with the input format of the training model.

[0038] S4. Generate preliminary prediction values ​​for dual objectives: Call two independently trained regression models, input the preprocessed prediction feature sets respectively, and generate preliminary prediction values ​​for the highest water level and the lowest water level corresponding to the prediction time node. Specifically, two independently trained random forest regression models are invoked ( , ), model parameters are set to .

[0039] If subsequent model validation results show that the root mean square error (RMSE) of the predicted values ​​is greater than 50% of the historical water level standard deviation, then... Adjust the value to 200, retrain the model, and generate predictions.

[0040] in, Using the historical highest water level as the training target, the preprocessed prediction feature set is input into the model, and then... Generate preliminary forecast values ​​of the highest water levels corresponding to the forecast time points ( ); Using the historical lowest water level as the training target, similarly through... Generate preliminary forecast values ​​of the lowest water levels corresponding to the forecast time points ( ).

[0041] S5. Verification, Correction, and Output: Logical verification is performed on the preliminary prediction values ​​of the dual targets. Predictions that do not conform to logic are corrected, and a preliminary prediction result containing key monitoring borehole information and the corrected prediction values ​​is output. Specifically, this includes: S51. Determine the preliminary predicted value of the lowest water level ( Is it greater than the preliminary forecast value of the highest water level? If so, then directly exchange the preliminary predicted values ​​of the lowest water level and the highest water level, and after correction, satisfy the condition that the preliminary predicted value of the lowest water level is less than or equal to the preliminary predicted value of the highest water level.

[0042] Judgment after correction and If the value exceeds ±3 times the standard deviation of the historical water level value of the borehole, it will be marked as an outlier and a manual verification will be prompted.

[0043] S52. Output preliminary prediction results including key information of the monitoring borehole, corrected dual-target prediction values, and preprocessing records. Key information of the monitoring borehole includes: unique borehole identifier, borehole latitude coordinates, borehole longitude coordinates, borehole elevation, and prediction time node; preprocessing records include: missing value imputation, normalized feature names, and processing parameters.

[0044] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-source fusion method for predicting groundwater levels, characterized in that, include: S1. Determine the prediction time node: Based on the multi-source groundwater monitoring data, extract the latest monitoring month as the base month, and calculate the next month after the base month as the prediction time node; S2. Construct a predictive feature set: For each monitored borehole, extract the dynamic monitoring data and borehole static parameters of the baseline month, and construct a predictive feature set that includes basic features and lag features. S3. Preprocessing the predicted feature set: Fill in the missing values ​​in the predicted feature set and normalize the numerical features to ensure that the feature set is compatible with the input format of the training model. S4. Generate preliminary prediction values ​​for dual objectives: Call two independently trained regression models, input the preprocessed prediction feature sets respectively, and generate preliminary prediction values ​​for the highest water level and the lowest water level corresponding to the prediction time node. S5. Verification, Correction and Output: Perform logical verification on the preliminary prediction values ​​of the dual targets, correct prediction values ​​that do not conform to logic, and output the preliminary prediction results that include key information of the monitoring borehole and the corrected prediction values.

2. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, The specific rules for calculating the predicted time nodes are as follows: If the base month is December, the prediction time point is the base year plus 1, and the month is fixed at 1; If the base month is from January to November, the prediction time node is the month with the base year unchanged and the month is the base month plus 1. The prediction time node is uniformly converted into a string in the format "YYYY-MM".

3. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, The static parameters of the borehole mentioned in step S2 include: unique borehole identifier, latitude, longitude, elevation, rock quality index and Lloyd's value; The dynamic monitoring data for the baseline month includes rainfall, highest water level, and lowest water level.

4. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, The basic features include: the monthly values ​​corresponding to the predicted time point, the borehole static parameters, and the rainfall in the baseline month; The lagging characteristics include rainfall, highest water level, and lowest water level for the base month.

5. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, The specific method for filling in missing values ​​in step S3 is as follows: If there are missing feature values ​​in the predicted feature set, the missing feature values ​​include: random missing values ​​or missing first values ​​of time series. The average value of the same type of feature values ​​of all other monitored boreholes in the same area is used to fill the missing values. If there are no other monitored boreholes in the area, the historical statistical mean of this feature is used to fill the feature, and the feature value is marked as a fill identifier in the feature set.

6. The multi-source fusion groundwater level prediction method according to claim 5, characterized in that, The specific information included in the filling identifier is: the name of the missing feature, the area range used for filling, the number of boreholes used in the mean calculation, and the filling time.

7. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, In step S3, the normalization process uses the Min-Max normalization method, specifically mapping the numerical features to the [0,1] interval. The mapping formula is as follows: in, These are the original eigenvalues. This is the minimum value of the feature. This is the maximum value of the feature, eliminating the impact of differences in the magnitude of different features on the model weight allocation.

8. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, In step S4, both independently trained regression models are random forest regression models, and the model parameters are set as follows: ; One model is trained on the historical highest water level to output a preliminary prediction of the highest water level; the other model is trained on the historical lowest water level to output a preliminary prediction of the lowest water level.

9. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, The specific process of logic verification in step S5 is as follows: Determine whether the preliminary forecast value of the lowest water level is greater than the preliminary forecast value of the highest water level. If so, directly swap the preliminary forecast values ​​of the lowest and highest water levels, and after correction, satisfy the condition that the preliminary forecast value of the lowest water level is less than or equal to the preliminary forecast value of the highest water level.

10. The multi-source fusion groundwater level prediction method according to claim 1, characterized in that, The key information monitored in step S5 includes: borehole unique identifier, borehole latitude coordinates, borehole longitude coordinates, borehole elevation, and predicted time node. The preliminary prediction results also include preprocessing records of the predicted feature set, which include: missing value imputation, normalized feature names and processing parameters.

Images