Integrated management and decision support system for river basin water environment based on big data analysis

Abstract

The application discloses a watershed water environment comprehensive management and decision support system based on big data analysis and relates to the technical field of water environment management.The system comprises a real-time acquisition module, a data regularization module, a self-adaptive fusion module, a risk diagnosis module and a decision optimization module; multi-source original data are collected through sensors; structured data sets are output after preprocessing; then, a multi-dimensional dynamic correlation fusion algorithm is used to realize deep fusion of time and space characteristics, generate a water environment state index, a trend prediction value and a key driving factor ranking; a dynamic risk assessment report is output based on a Bayesian enhanced integrated model and a self-adaptive threshold mechanism; and an intervention scheme with optimal comprehensive benefits is generated through a multi-objective optimization model.The application realizes intelligent processing of watershed water environment data in the whole chain, improves the accuracy of environmental state assessment, the timeliness of risk prediction and the scientificity of decision schemes, and provides efficient technical support for comprehensive management of watershed water environment.

Description

Technical Field

[0001] This invention relates to the field of water environment management technology, and more specifically, to a watershed water environment integrated management and decision support system based on big data analysis. Background Technology

[0002] Existing watershed water environment management systems primarily focus on basic monitoring and simple data presentation. They acquire core water quality parameters by deploying conventional monitoring equipment and then utilize basic statistical analysis and visualization tools to provide management departments with basic water environment status information, playing a certain role in initial water environment supervision and preliminary investigation.

[0003] However, as watershed water environment issues become increasingly complex, the limitations of such systems are becoming increasingly apparent: They fail to comprehensively capture water environment influencing factors, focusing primarily on core water quality indicators while lacking sufficient integration of related data such as hydrological and meteorological data. Furthermore, their sampling strategies lack flexibility, making it difficult to cope with sudden changes and complex operating conditions in the water environment. Data preprocessing methods are relatively basic, failing to adequately address outliers and missing values ​​in monitoring data, and neglecting to fully consider the spatiotemporal correlations of the data, resulting in data quality that cannot meet the needs of in-depth analysis. Data analysis often remains at the level of surface statistics and simple integration, failing to uncover the complex relationships behind the data. The quantitative assessment of the overall state of the water environment is not comprehensive enough, and the accuracy and foresight of trend predictions are insufficient. Risk assessment relies heavily on fixed standards, failing to fully adapt to the dynamic changes in the water environment, and the practicality of the assessment results needs improvement. In terms of decision-making recommendations, they often fail to consider multiple practical needs, resulting in insufficient comprehensiveness and operability of the solutions, making it difficult to support refined and scientific management decisions. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a comprehensive watershed water environment management and decision support system based on big data analysis. The system addresses the problems mentioned in the background section of the prior art, such as difficulty in uncovering deep dynamic correlations, lack of accurate prediction and comprehensive quantification, reliance on fixed thresholds for risk assessment, and a focus on a single objective in decision-making.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a watershed water environment integrated management and decision support system based on big data analysis, comprising: a real-time acquisition module: used to collect multi-source raw data within the target watershed to form a multi-source raw dataset; The multi-source raw data includes: pH value, dissolved oxygen, ammonia nitrogen concentration, water flow rate, rainfall, and water temperature; Data normalization module: preprocesses the original multi-source time series dataset and outputs the preprocessed structured dataset; The preprocessing includes outlier detection and removal, missing value imputation, temporal alignment, spatial gridding, and standardization. Adaptive fusion module: Based on the structured dataset, a multidimensional dynamic association fusion algorithm is used to perform deep fusion and dynamic association mining of spatiotemporal multidimensional features, and output a fusion analysis result set; The fusion analysis results set includes a comprehensive water environment status index, short-term trend forecasts, and ranking of key driving factors. Risk Diagnosis Module: Receives the fusion analysis result set, adopts uncertainty quantification and adaptive threshold adjustment mechanism, and based on Bayesian enhanced machine learning ensemble model, sorts and maps the comprehensive water environment state index, short-term trend prediction value and key driving factors to dynamic risk level, and outputs risk assessment report set; Decision optimization module: Based on the risk assessment report set, construct a multi-objective optimization model to simulate and generate intervention plans, calculate the comprehensive benefit score of the plans, rank and recommend the optimal plan and alternative plans, and output decision support results.

[0006] The technical effects and advantages of this invention are as follows: 1. This invention establishes a multi-source data collaborative acquisition system, relying on high-precision dedicated sensors to collect core parameters related to water quality, hydrology, and meteorology. It adopts a dynamically adaptable sampling frequency strategy and a standardized data transmission protocol to ensure real-time data upload. This achieves comprehensive coverage and accurate capture of key influencing factors of the watershed's water environment, effectively improving the integrity and timeliness of the original data and laying a solid data foundation for subsequent full-chain analysis. 2. This invention systematically streamlines the entire preprocessing logic, including outlier detection, missing value imputation, time alignment, spatial gridding, and standardization. It combines statistical methods with spatiotemporal correlation characteristics to refine data processing. This not only solves problems such as anomalies, missing values, and inconsistent time sequences in the original data, but also achieves spatiotemporal unification and dimensionlessness of multi-source data. This results in a structured dataset with good consistency, comparability, and reliability, providing high-quality data support for in-depth fusion analysis. 3. This invention employs a three-dimensional convolutional neural network combined with an attention mechanism for deep fusion of spatiotemporal features, leverages graph neural networks to mine dynamic correlations between parameters, and combines principal component analysis, entropy weight method, long short-term memory network and SHAP value analysis to not only generate a comprehensive index that can fully quantify the state of the water environment, but also accurately complete short-term trend prediction and ranking of key driving factors, achieving in-depth analysis of water environment data and significantly improving the scientificity and accuracy of the analysis results. 4. This invention assesses the reliability of core analysis results by introducing an uncertainty quantification mechanism, dynamically adapts to changes in the water environment by combining an adaptive threshold adjustment mechanism, and maps risk levels using a Bayesian-enhanced machine learning ensemble model. This breaks the limitations of traditional fixed threshold assessment, making risk level classification more in line with actual conditions. At the same time, the rigor of the assessment is ensured by probability confidence and manual review prompts, which significantly improves the reliability and dynamic adaptability of risk assessment. 5. This invention constructs a multi-objective optimization model covering economic, environmental, and social dimensions, combines it with a full-chain intervention measure library, uses a dedicated algorithm to solve and screen the Pareto optimal solution set, and ranks and recommends the optimal solution and classified alternative solutions through a three-level system comprehensive benefit scoring and scenario simulation robustness test. It also provides detailed implementation guidelines and dynamic adjustment suggestions, providing comprehensive, feasible, and optimal decision support for watershed water environment management, effectively improving the refinement and scientific level of management work. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the overall system structure of the present invention; Figure 2 This is a schematic diagram of the data preprocessing process of the present invention; Figure 3 This is a schematic diagram of the process for obtaining the fusion analysis result set of the present invention; Figure 4 This is a schematic diagram illustrating the process of obtaining the risk assessment report set according to the present invention; Figure 5 This is a schematic diagram illustrating the changes in the overall benefit score of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] As attached Figures 1 to 5 The integrated watershed water environment management and decision support system based on big data analysis shown includes: a real-time acquisition module: used to collect multi-source raw data within the target watershed to form a multi-source raw dataset; The multi-source raw data includes: pH value, dissolved oxygen, ammonia nitrogen concentration, water flow rate, rainfall, and water temperature; The process of collecting multi-source raw data within the target watershed to form a multi-source raw dataset is as follows: pH value acquisition: HACH pH online pH electrode is selected, with a measurement range of 0 to 14 and an accuracy of ±0.01. It adopts the principle of glass electrode combined with reference electrode and is equipped with automatic temperature compensation function. The sensor is deployed at a position 0.5m below the water surface at key sections of the watershed. It is equipped with an automatic cleaning device, which uses high-pressure air to regularly remove dirt from the electrode surface. The sampling frequency is once every 5 to 15 minutes, which can be dynamically adjusted to once every 1 minute during pollution events. Data is output through RS485 / Modbus protocol.

[0010] Dissolved oxygen acquisition: A HACHLDO fluorescence dissolved oxygen sensor is used, with a measurement range of 0 to 20 mg / L and an accuracy of ±0.1 mg / L. It adopts the fluorescence quenching principle, eliminating the need for frequent calibration and membrane replacement. The sensor is deployed in the same location as the pH sensor, fixed on the same monitoring platform, and equipped with an automatic wiping and cleaning device to prevent biofilm adhesion from affecting measurement accuracy. The sampling frequency is synchronized with the pH value, once every 5 to 15 minutes, and the data is output via RS485 / Modbus protocol.

[0011] Ammonia nitrogen concentration acquisition: The HACHAN-ISE ultraviolet absorption ammonia nitrogen sensor is selected, with a measurement range of 0 to 100 mg / L and an accuracy of ±2%. It requires no reagents, has a long maintenance cycle, and the sensors are deployed at the same monitoring section at a depth of 0.5 m below the water surface. It is equipped with a filtration and automatic cleaning system to avoid interference from suspended solids. The sampling frequency is synchronized with pH value and dissolved oxygen. Data is output via RS485 / Modbus protocol.

[0012] Water temperature acquisition: A PT100 platinum resistance temperature sensor is selected, with a measurement range of -5 to 50℃ and an accuracy of ±0.1℃. The sensor is integrated into the water quality monitoring platform, directly contacting the water body, with a response time of less than 30 seconds. The sampling frequency is fully synchronized with other water quality parameters, and the data is output uniformly through the RS485 / Modbus protocol along with the platform.

[0013] Water flow data acquisition: HACHFLO-DAR radar flow meter is selected, with a measurement range of 0 to 10 m / s and an accuracy of ±2%. Non-contact radar measurement is used to avoid clogging of underwater components. The flow meter is fixed at a high position on the bridge or bank and installed in a stable section of the river cross-section. The water level-flow relationship curve method is used for auxiliary calibration. Water level data is collected synchronously using an ultrasonic water level meter with an accuracy of ±3 mm. The sampling frequency is once every 5 to 10 minutes, and it is triggered synchronously with water quality parameters to ensure the consistency of hydrological and water quality data time sequence.

[0014] Rainfall data acquisition: Davis Vantage Pro2 tipping bucket rain gauges with a resolution of 0.2 mm and a measurement accuracy of ±5% were selected. The rain gauges were deployed at representative meteorological stations within the watershed, covering at least one station each in the upstream, midstream, and downstream sections. The installation height was 1.5 m above the ground with no obstructions. At the same time, real-time rainfall grid data from the regional meteorological department was connected to the data, with a resolution of 1 km × 1 km, as a supplement and for verification. The sampling frequency was once every 1 to 5 minutes, and automatically increased to once every 30 seconds during periods of high-intensity rainfall.

[0015] Data normalization module: preprocesses the original multi-source time series dataset and outputs the preprocessed structured dataset; The preprocessing includes outlier detection and removal, missing value imputation, temporal alignment, spatial gridding, and standardization. The original multi-source time series dataset is preprocessed as follows: Outlier Detection and Removal: Outlier identification is performed on the original multi-source time-series dataset based on statistical methods and physical thresholds. First, the three-standard-deviation principle is applied; for each parameter's time series, the mean μ and standard deviation σ of the series are calculated. If a data point x... i Satisfy |x i If -μ∣>3σ, it is marked as abnormal; secondly, the physical reasonable range of each parameter is set, for example, the pH value is limited to between 4 and 10, and the dissolved oxygen is limited to between 0 and 20 mg / L. Data points that exceed this range are directly marked as abnormal; at the same time, a moving window mutation detection is adopted, with a window length of 30 sampling points. If the change rate of adjacent data points exceeds the preset threshold, such as the change rate of water flow rate being greater than 20%, it is marked as abnormal; all data points marked as abnormal are removed, and their location, parameter type, and marking reason are recorded for easy follow-up.

[0016] Missing value imputation: For missing data caused by outlier removal or transmission interruption, time series imputation methods are used; for cases with fewer than 5 consecutive missing points, linear imputation is used, with the formula as follows: Where t is the missing time, k is the number of consecutively missing points in the segment, and x t-1 x is the nearest valid value before the missing segment. t+k x is the nearest valid value after the missing segment. t The result is the imputation result at the missing time t. For long, continuous missing sequences, the nearest neighbor imputation method is used to select the 5 most recent effective neighbor points in time and fill the missing values ​​by Euclidean distance weighted average. For parameters with obvious periodicity, such as rainfall, seasonal imputation is performed by combining the historical average value for the same period. After imputation, the imputation segment is smoothed to ensure the overall continuity of the sequence.

[0017] Time alignment: Unify multi-source data with different sampling frequencies to a standard time grid; use 1 minute as the base time step, downsample and average high-frequency data (sampling frequency < 1 minute); for low-frequency data (sampling frequency > 1 minute), first upsample to 1 minute granularity through linear interpolation, and then use the nearest neighbor resampling method to correct timestamp deviation; for data with slight time offset, adjust to a unified timestamp through linear interpolation to ensure that all parameters have corresponding values ​​at the same time point, thereby forming a multi-source time series table with strict time consistency.

[0018] Spatial gridding: Point source data distributed across different monitoring stations are mapped to a unified watershed spatial grid; the grid resolution is set to 1 km × 1 km, and ordinary kriging interpolation is used to calculate spatial weights based on the latitude and longitude coordinates and variograms of the monitoring stations; for each grid center point, its parameter value is obtained by weighted averaging of surrounding stations, and the weight formula is as follows: Where s0 is the center point of the grid, z(s) i ) represents the measured value of the i-th monitoring station, and λ i For the corresponding kriging weights, The interpolated values ​​of the parameters for the watershed grid center point s0 are given. These weights satisfy the conditions of unbiasedness and minimum variance. Considering the influence of water flow direction and topography, the weights are adjusted for anisotropy, and finally, a spatiotemporal gridded dataset covering the entire watershed is output.

[0019] Standardization: All parameters are dimensionless; zero-mean standardization is used to calculate the mean μ and standard deviation σ for each parameter series. The standardization formula is as follows: Where x i z is the original value. i The value is the standardized value; For parameters with strong skewed distributions, such as ammonia nitrogen concentration, logarithmic transformation is performed first, followed by standardization. After standardization, the values ​​of each parameter are basically distributed between -3 and 3, ensuring that parameters of different dimensions are comparable in subsequent fusion analysis.

[0020] It should be further explained that the above method enables comprehensive preprocessing of the original multi-source time series dataset, outputting a high-quality structured dataset.

[0021] Adaptive fusion module: Based on the structured dataset, a multidimensional dynamic association fusion algorithm is used to perform deep fusion and dynamic association mining of spatiotemporal multidimensional features, and output a fusion analysis result set; The fusion analysis results set includes a comprehensive water environment status index, short-term trend forecasts, and ranking of key driving factors. Based on the structured dataset, a multidimensional dynamic association fusion algorithm is used to perform deep fusion and dynamic association mining of spatiotemporal multidimensional features. The specific process is as follows: Spatiotemporal multidimensional feature deep fusion: First, spatiotemporal convolution operations are performed on the structured dataset to capture the spatial dependencies and temporal evolution relationships between parameters. A three-dimensional convolutional neural network model is used to extract features from the gridded data tensor. The convolution kernel size is set to 3×3×3, corresponding to the time, latitude, and longitude dimensions, respectively. Through multi-layer convolution and pooling operations, parameters such as pH value, dissolved oxygen, ammonia nitrogen concentration, water temperature, water flow, and rainfall are fused to form a high-dimensional feature vector. Specifically, the first convolution layer uses 32 filters with a stride of 1 and the activation function is a modified linear unit. The second convolution layer uses 64 filters with a stride of 2 and combines max pooling to reduce dimensionality. An attention mechanism is introduced during the fusion process to dynamically allocate weights to key spatiotemporal regions. For example, the weight of water quality parameters in high-pollution sections is increased by 20%, thereby generating the fused spatiotemporal feature map.

[0022] Dynamic correlation mining: Based on the fused spatiotemporal feature map, a graph neural network model is used to mine the dynamic correlations between parameters; monitoring stations are regarded as graph nodes, and the correlation between parameters is regarded as edge weights. The initial edge weights are calculated using the Pearson correlation coefficient, and the formula is as follows: Where n is the sequence length, x i and y i For the i-th corresponding value of the two parameter sequences, and The mean values ​​of each node are used; a graph attention network is employed to propagate and update node features, iterating through 3 to 5 layers, with 4 attention heads per layer, to capture the immediate impact of rainfall on water flow or the delayed correlation between ammonia nitrogen concentration and dissolved oxygen. Specifically, this includes calculating attention coefficients. The formula is:

[0023] Where h i and h j Let W be the node feature vector, W be the linear transformation matrix, a be the attention parameter vector, and N be the node feature vector. i For the neighborhood of node i, the mining results are output in the form of an association matrix.

[0024] It should be further explained that the fusion analysis result set is obtained in the following way: Calculation of the comprehensive water environment state index: The dynamic correlation mining results are input into the weighted fusion model to calculate the comprehensive water environment state index. Principal component analysis is used to reduce the dimensionality of multidimensional features and retain the principal components with a cumulative variance contribution rate of 85%. Specific steps include calculating the covariance matrix, solving for eigenvalues ​​and eigenvectors, selecting the first few principal components, and then determining the weights of each parameter using the entropy weight method. The formula is as follows: Where m is the number of samples, p ijThe standardized probability of the j-th parameter for the i-th sample (calculated by first standardizing the j-th parameter to its zero mean, z). ij Perform nonnegation processing Then calculate the proportion. ), p is the total number of parameters, e j Let the information entropy of the j-th parameter be , and the final exponential formula be . , where f j The standardized value of the j-th parameter is given by an index ranging from 0 to 100. A higher value indicates a better water environment. The correlation matrix is ​​considered to adjust the weights during the calculation to ensure that strongly correlated parameters have higher weights. Some experimental data are shown in the table below:

[0025] Short-term trend prediction generation: Based on fused features and the correlation matrix, a Long Short-Term Memory (LSTM) network model is used for short-term trend prediction. The input sequence length is set to the past 24 hours of data, and the prediction window is the next 6 to 12 hours. The model contains two layers of LTM units, with 128 hidden units in each layer. The mean squared error loss function is used for training, and the formula is as follows: Where L is the model training loss value, q is the number of prediction points, and y is the number of prediction points. t For the true value, For the predicted values, the parameters are iteratively optimized using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The prediction output includes the trend sequence of each parameter and the predicted value of the overall water environment index. To improve robustness, the Monte Carlo dropout method is integrated to simulate uncertainty.

[0026] Key driving factors ranking: The SHAP value method is used to quantify the contribution of each parameter to the comprehensive water environment state index; based on the gradient boosting tree surrogate model, the SHAP value of each input parameter is calculated, and the formula involves the expected value difference. Where N is the set of all parameters, S is the subset excluding the j-th parameter, v(·) is the model output function, and φ j The SHAP value of the j-th parameter; ! represents factorial operation, for example, |S|! is the factorial of the number of elements in set S; sort by the absolute value of the SHAP value in descending order, output the top 5 key driving factors, such as rainfall or ammonia nitrogen concentration, with the percentage of contribution. In the specific implementation, TreeExplainer from the SHAP library is used to interpret the surrogate model to ensure consistent interpretation.

[0027] It should be further explained that, through the above method, in-depth analysis and mining of structured datasets are achieved, and a fusion analysis result set including comprehensive water environment state index, short-term trend forecast values, and ranking of key driving factors is output.

[0028] Risk Diagnosis Module: Receives the fusion analysis result set, adopts uncertainty quantification and adaptive threshold adjustment mechanism, and based on Bayesian enhanced machine learning ensemble model, sorts and maps the comprehensive water environment state index, short-term trend prediction value and key driving factors to dynamic risk level, and outputs risk assessment report set; The specific process of the uncertainty quantification and adaptive threshold adjustment mechanism is as follows: Uncertainty Quantification: First, the key outputs of the fusion analysis result set are subjected to uncertainty assessment to quantify their reliability. For the comprehensive water environment state index, the Monte Carlo simulation method was used. Gaussian noise (mean 0, standard deviation 5% of the index value) was added to the input features, and 1000 forward propagations were performed to calculate the empirical distribution of the index, obtaining a 95% confidence interval. The formula is as follows: ,in The index is the average of 1000 simulations. CI95% represents the 95% confidence interval of the comprehensive water environment state index. σ I The standard deviation is N=1000; For short-term trend predictions, posterior distribution sampling is used with a Bayesian Long Short-Term Memory network, combined with Monte Carlo dropout (dropout rate 0.2) enabled during the inference phase to generate probability bands for the predicted values. The width of these bands is calculated using the formula W=q. 97.5 -q 2.5 Where q 97.5 and q 2.5 These are the 97.5% and 2.5% quantiles of the predicted distribution, respectively, and W represents the width of the predicted probability band. For the ranking of key driving factors, the standard error of the SHAP value is calculated, and the confidence interval of the contribution percentage is estimated by resampling 100 times using the guided method. Finally, the three types of uncertainty are normalized to the range of 0 to 1 and merged into a comprehensive uncertainty score u, which is then weighted and averaged (exponential weight 0.4, prediction weight 0.4, driving factor weight 0.2).

[0029] Adaptive threshold adjustment mechanism: Risk level thresholds are dynamically adjusted based on uncertainty scores. Initial risk thresholds are set as low risk (comprehensive water environment state index 0 to 30), medium risk (31 to 60), and high risk (61 to 100). The adjustment rule uses fuzzy logic: if the uncertainty score u < 0.3, the initial threshold is maintained; if 0.3 ≤ u < 0.6, the low-risk threshold shifts conservatively by 4.5%-9%, and the medium-risk threshold shifts by 9%-18% (the shift ratio increases linearly with u); if u ≥ 0.6, the low-risk threshold shifts by 9%-15%, and the medium-risk threshold shifts by 18%-30% (the shift ratio increases linearly with u). The specific adjustment formula is as follows: Where β=0.3 is the adjusted strength coefficient, Tlow =61 (lower risk limit), T mid =31 (lower limit of medium risk), T high =0 (lower limit of high risk) These are the adjusted lower threshold for low risk and the adjusted lower threshold for medium risk, respectively, where u represents the comprehensive uncertainty score (0~1). The influence of key driving factors is also introduced; if the top three driving factors include highly sensitive parameters (such as ammonia nitrogen concentration or dissolved oxygen), the medium-to-high risk boundary is tightened by an additional 5%. The threshold adjustment is recalculated every 30 minutes to ensure adaptation to rapid changes in the aquatic environment.

[0030] It should be further explained that the Bayesian-enhanced machine learning ensemble model is constructed as follows: The input feature vector (dimension 8) is composed of the current value of the comprehensive water environment state index, the mean and slope of the short-term trend prediction, and the contribution percentage of the top five key driving factors. The basic ensemble model consists of three parts: a random forest (200 trees, maximum depth 15), a gradient boosting tree (150 trees, learning rate 0.05), and an additional lightweight gradient boosting machine (100 trees). The class probabilities output by these three parts are fused through a multi-class logistic regression (Softmax) stacked layer. Bayesian augmentation uses variational inference to approximate the posterior distribution, with the prior set to a normal distribution and the variational family to a mean Gaussian distribution. The lower bound of evidence (ELBO) is optimized, and the formula is as follows: Where q(θ) is the variational posterior, p(θ) is the prior, D is the historical labeled risk data (containing 10,000 historical samples of the watershed), KL(·||·): KL divergence (measures the difference between two distributions), and θ is the vector of parameters to be estimated in the Bayesian model; Adam is used for optimization, and after 200 iterations, the parameter posterior samples are obtained and used to generate the risk probability distribution in the inference stage.

[0031] It should be further explained that the mapping process for the dynamic risk level is as follows: The feature vectors are input into the Bayesian augmented ensemble model to obtain the posterior probability distributions for each risk level (low, medium, and high). The risk level is initially determined using the maximum a posteriori probability principle. If the maximum probability is less than 0.7, uncertainty correction is triggered: combining an adaptive threshold and uncertainty score, if the comprehensive water environment state index is lower than the adjusted medium-risk threshold and the uncertainty score is greater than 0.5, a mandatory upgrade to one level is performed. Simultaneously, the evidence ratio supporting the current level is calculated using the following formula: ER is the evidence ratio (the confidence level used to determine the risk level). If ER < 2, it is marked as "requires manual review". The final output is the dynamic risk level and the corresponding probability confidence level.

[0032] Finally, all the intermediate results are integrated to form a set of structured risk assessment reports.

[0033] Decision optimization module: Based on the risk assessment report set, construct a multi-objective optimization model to simulate and generate intervention plans, calculate the comprehensive benefit score of the plans, rank and recommend the optimal plan and alternative plans, and output decision support results.

[0034] The multi-objective optimization model is constructed as follows: The multi-objective optimization model takes minimizing economic costs, maximizing environmental benefits, and minimizing social impact as its three core objectives. A nonlinear multi-objective optimization function is constructed, and each objective is normalized using 0-1 to eliminate dimensional differences, as detailed below: The objective formula for minimizing economic costs is: C con C is the one-time construction cost; op (t) represents the operation and maintenance cost in year t; r = 3.5% is the benchmark discount rate, t is the year index; T is the service life (1 year for emergency measures, 8-12 years for engineering measures, and 5 years for management measures); C res Residual value (10%-20% for engineering projects, 5% for management projects); C max This represents the highest cost among similar historical solutions. The formula for maximizing environmental benefits is: , where I pre To determine the comprehensive water environment status index before intervention, I post I is the predictive index after intervention. max =100; F i,pre F i,post The values ​​before and after intervention for the top 5 key driving factors; w I =0.5, w i For the driving factor weights (calculated using the entropy weight method), ); shap i Let be the SHAP value of the i-th key driver, sign(-shap) i ) is the sign coefficient (for the negative factor shap) i <0 is taken as +1, for positive factors shap i >0 is taken as +1), ensuring that the positive improvement of all driving factors is reflected in the improvement of environmental benefits; The formula for minimizing social impact is: ; where L prod L life Impact on residents' lives (on a scale of 1-5, with higher scores indicating greater impact), L eco Ecological disturbance (the larger the value, the more severe the disturbance); L prod,max L life,max L eco,max This is the historical maximum value; w prod =0.4、wlife = 0.3, w eco = 0.3; The constraint conditions include: resource constraints (funds within ±20% of the annual budget, professional staff ≤ 30% of the existing, emergency equipment ≤ 80% of the reserves); technical feasibility constraints (ammonia nitrogen removal rate ≥ 30%, pollutant interception rate ≥ 25%, emergency plan effective within 6 - 12 hours); environmental carrying capacity constraints (ammonia nitrogen ≤ 1.0 mg / L, dissolved oxygen ≥ 5.0 mg / L, index compliance or improvement ≥ 20%); time - sequence coordination constraints (start emergency within 4 hours when risk escalates within 8 hours, start the project within 3 months for medium - and long - term risks).

[0035] The calculation method of the comprehensive benefit score of the said plan is as follows: The comprehensive benefit score adopts a three - level system of target layer - criterion layer - index layer, determines the weights by the entropy weight method, and the comprehensive benefit S of the target layer takes values from 0 to 100 points; The criterion layer includes economic rationality (weight 0.3), environmental improvement degree (weight 0.5), and social acceptability (weight 0.2); in the index layer, economic rationality includes the cost of unit environmental benefit and the cost recovery period (both are negative indicators), environmental improvement degree includes the index improvement rate, the compliance rate of driving factors, and the decline range of risk level (all are positive indicators), and social acceptability includes the satisfaction of residents (positive), the industrial loss rate (negative), and the ecological disturbance restoration period (negative); Standardization formula: positive indicator negative indicator where x m,n is the measured value of the indicator, xm,n,max and xm,n,min are the extreme values of this indicator in the candidate plans, m is the index of the criterion layer, and n is the index of the index layer; The scoring formula is where w m is the weight of the criterion layer, k m is the number of indicators in the m - th criterion layer, w m,n is the indicator weight ( ); The robustness test calculates the coefficient of variation through scenario simulation (rainfall ±20%, treatment cost ±15%), where is the standard deviation of the score, is the average score of the score; CV ≤ 0.15 indicates that the robustness meets the standard, 0.15 < CV ≤ 0.3 requires additional scenario adaptation explanations, and CV > 0.3 directly eliminates this plan.

[0036] The acquisition method of the said decision - making support result is as follows: The decision - making support result is obtained through three links: intervention plan generation, plan sorting and screening, and structured output, specifically as follows: Intervention plans are generated based on dynamic risk levels and key driving factors, constructing a full-chain intervention measure library: Source control and emission reduction measures are applicable to medium-to-high risk scenarios dominated by ammonia nitrogen concentration, affecting ammonia nitrogen concentration and pollutant emission intensity; engineering treatment measures are applicable to medium-risk scenarios of continuous water quality deterioration, affecting ammonia nitrogen concentration, dissolved oxygen, and water temperature; hydrological scheduling measures are applicable to high-risk scenarios dominated by water flow or rainfall, affecting water flow, rainfall, and pollutant dilution ratio; emergency response measures are applicable to high-risk sudden pollution events, affecting ammonia nitrogen concentration and dissolved oxygen; and management and control measures are applicable to low-to-medium risk scenarios with potential escalation, affecting all driving factors (risk warning and prevention). A non-dominated sorting genetic algorithm (NSGA-III) is used to solve the problem, with mixed encoding representing whether measures are implemented (binary: 0 = not implemented, 1 = implemented) and the intensity of implementation (real numbers: e.g., production restriction ratio 0.2-0.8, water diversion volume 100,000-500,000 m³). 3 (per day), with a population size of 100, number of iterations of 200, crossover probability of 0.8, and mutation probability of 0.05, a Pareto optimal solution set is selected by combining an elite retention strategy to form 20-30 candidate solutions; The schemes are ranked according to the following priority rules: comprehensive benefit score in descending order, robustness coefficient of variation in ascending order, and implementation period in descending order. The optimal scheme is the top-scoring scheme that meets the robustness standard, and must meet the following requirements: environmental improvement score ≥ 40 points (out of 50 points) and social acceptability score ≥ 15 points (out of 20 points). The alternative schemes are the second to fourth-scoring schemes that meet the robustness standard, namely, the economically optimal scheme (economic rationality score ≥ 25 points, out of 30 points), the emergency-efficient scheme (implementation period ≤ 72 hours), and the long-term stable scheme (service life ≥ 10 years, CV ≤ 0.1). The structured output includes details of the optimal solution (list of measures, implementation steps, resource allocation, and expected results), a comparison of core indicators of alternative solutions, implementation risk warnings (uncertainties and contingency plans), and dynamic adjustment suggestions (updated every 6 hours based on the latest data).

[0037] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A comprehensive watershed water environment management and decision support system based on big data analysis, characterized in that: include: Real-time acquisition module: used to collect multi-source raw data within the target watershed to form a multi-source raw dataset; The multi-source raw data includes: pH value, dissolved oxygen, ammonia nitrogen concentration, water flow rate, rainfall, and water temperature; Data shaping module: preprocesses the multi-source raw dataset and outputs the preprocessed structured dataset; The preprocessing includes outlier detection and removal, missing value imputation, temporal alignment, spatial gridding, and standardization. Adaptive fusion module: Based on the structured dataset, a multidimensional dynamic association fusion algorithm is used to perform deep fusion and dynamic association mining of spatiotemporal multidimensional features, and output a fusion analysis result set; The fusion analysis results set includes a comprehensive water environment status index, short-term trend forecasts, and ranking of key driving factors. Risk Diagnosis Module: Receives the fusion analysis result set, adopts uncertainty quantification and adaptive threshold adjustment mechanism, and based on Bayesian enhanced machine learning ensemble model, sorts and maps the comprehensive water environment state index, short-term trend prediction value and key driving factors to dynamic risk level, and outputs risk assessment report set; Decision optimization module: Based on the risk assessment report set, construct a multi-objective optimization model to simulate and generate intervention plans, calculate the comprehensive benefit score of the plans, rank and recommend the optimal plan and alternative plans, and output decision support results.

2. The watershed water environment integrated management and decision support system based on big data analysis according to claim 1, characterized in that: The methods for collecting the multi-source raw data are as follows: pH value: An online pH electrode was used, with a measurement range of 0 to 14, an accuracy of ±0.01, and a sampling frequency of once every 5 to 15 minutes; Dissolved oxygen: A HACH LDO fluorescence dissolved oxygen sensor was selected, with a measurement range of 0 to 20 mg / L, an accuracy of ±0.1 mg / L, and a sampling frequency of once every 5 to 15 minutes; Ammonia nitrogen concentration: The HACH AN-ISE ultraviolet absorption ammonia nitrogen sensor was selected, with a measurement range of 0 to 100 mg / L, an accuracy of ±2%, and a sampling frequency of once every 5 to 15 minutes; Water flow rate: HACH FLO-DAR radar flow meter is selected, with a measurement range of 0 to 10 m / s and an accuracy of ±2%. It adopts non-contact radar measurement and the sampling frequency is once every 5 to 10 minutes. Rainfall: A Davis Vantage Pro2 tipping bucket rain gauge was used, with a resolution of 0.2 mm, a measurement accuracy of ±5%, and a sampling frequency of once every 1 to 5 minutes; Water temperature: A PT100 platinum resistance temperature sensor is used, with a measurement range of -5 to 50℃, an accuracy of ±0.1℃, and a sampling frequency of once every 5 to 15 minutes.

3. The integrated watershed water environment management and decision support system based on big data analysis according to claim 1, characterized in that: The preprocessing process is as follows: Outlier detection: Outliers are identified by combining the three-standard-deviation principle, physical reasonable range limitation, and moving window abrupt change detection. After removal, relevant information is recorded. Missing value imputation: Linear imputation is used for short sequence missing values, and nearest neighbor imputation is used for long sequence missing values. Periodic parameters are combined with historical average values ​​for imputation, and smoothing is performed after imputation. Time alignment: Using a one-minute time step, timestamps from multiple data sources are unified through downsampling, upsampling, and interpolation adjustments. Spatial gridding: Kriging interpolation is used to map point source data to a watershed spatial grid, and the weights are adjusted in combination with water flow direction and topography; Standardization process: Zero-mean standardization is used, and skewed distribution parameters are first logarithmically transformed and then standardized.

4. The integrated watershed water environment management and decision support system based on big data analysis according to claim 1, characterized in that: The process of deep fusion of spatiotemporal multidimensional features is as follows: spatiotemporal convolution operation is performed on the structured dataset, a three-dimensional convolutional neural network model is used to extract gridded data tensor features, and the parameters are fused through multi-layer convolution and pooling operations to form a high-dimensional feature vector. An attention mechanism is introduced to dynamically allocate the weights of key spatiotemporal regions and generate the fused spatiotemporal feature map.

5. The integrated watershed water environment management and decision support system based on big data analysis according to claim 1, characterized in that: The process of dynamic association mining is as follows: The monitoring stations are treated as graph nodes, and the parameter correlations are treated as edge weights. The initial edge weights are calculated using the Pearson correlation coefficient. A graph attention network is used to iteratively update the node features, capture the immediate impact and delayed correlation between parameters, calculate the normalized attention coefficients, and output the mining results in the form of an association matrix.

6. The integrated watershed water environment management and decision support system based on big data analysis according to claim 1, characterized in that: The process of quantifying uncertainty is as follows: Monte Carlo simulation was used to calculate the specified confidence interval for the comprehensive water environment state index; For short-term trend predictions, probability bands are generated by sampling the posterior distribution of a Bayesian long short-term memory network combined with Monte Carlo dropout. Confidence intervals for contribution percentages were estimated by guided resampling after ranking key driving factors. After normalizing the uncertainties of the three factors, they are weighted and merged according to the set weights to form a comprehensive uncertainty score.

7. The integrated watershed water environment management and decision support system based on big data analysis according to claim 1, characterized in that: The machine learning ensemble model is constructed as follows: The input feature vectors are the current value of the comprehensive water environment state index, the mean and slope of the short-term trend prediction, and the contribution percentage of the top five key driving factors. The basic model consists of random forest, gradient boosting tree and lightweight gradient boosting machine, and the output probability is fused by multi-class logistic regression stacking layers. Variational inference is used to optimize the lower bound of evidence, with normal distribution as the prior and Gaussian distribution of mean field as the variational family, and the parameter posterior sample is obtained by iterative optimization.

8. The integrated watershed water environment management and decision support system based on big data analysis according to claim 1, characterized in that: The calculation method for the comprehensive benefit score is as follows: a three-level system of target layer - criterion layer - indicator layer is adopted, and the weight of each level is determined by the entropy weight method. The target layer is the comprehensive benefit; the criterion layer includes economic rationality, environmental improvement degree and social acceptability; the indicator layer is divided into positive and negative indicators, which are standardized according to the corresponding formulas. The comprehensive benefit score is obtained by hierarchical weighted calculation, and the robustness test is carried out by calculating the coefficient of variation through scenario simulation. The degree of robustness compliance of the scheme is judged according to the size of the coefficient of variation.

Images