Water quality short-term prediction method based on parameter perturbation-localization ensemble data assimilation

Abstract

The application provides a water quality short-term prediction method based on parameter perturbation-localized ensemble data assimilation, belongs to the technical field of water quality prediction, and is characterized in that: based on the water quality state and uncertainty of a hydrodynamic-water quality mechanism model at an initial time T1, an initial state vector is constructed and converted into an initial ensemble member to input the model; sensitive water quality parameters of the model are screened, the parameters are perturbed, and the parameter perturbation set is generated together with the model driving conditions, and then the parameter perturbation set is fused with the initial ensemble member to form a perturbation set; the perturbation set is input into the model to run to the T2 time, and a water quality prediction set is obtained; the observation data and error of the water quality at the T2 time are combined, and an EnKF or other ensemble data assimilation algorithm is used to calculate and analyze the set; finally, the analysis set is used to update the initial conditions of the model at the T2 time, the above steps are repeated, and the optimal estimation at each time is continuously output, so that the water quality short-term prediction is realized. The application can improve the water quality short-term prediction precision, save the calculation cost, accurately capture the spatiotemporal heterogeneity of water bloom, and has strong interpretability and generalizability.

Description

Technical Field

[0001] This invention relates to the field of water quality forecasting technology, and in particular to a short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation. Background Technology

[0002] From the perspective of current technological trends, eutrophication and algal blooms have become major challenges in the field of water environment protection, with particularly significant impacts on the ecological security and water quality management of static water bodies such as lakes and reservoirs. Against this backdrop, short-term algal bloom forecasting, as a key technology for early prevention and control of algal bloom pollution and supporting refined water quality management, is evolving from early single-model simulations to a direction of fusion and optimization of observational data and model simulations. Among these, data assimilation methods, due to their ability to integrate observational results from different sources and at different resolutions with model simulation data to generate optimal water quality estimates that are consistent in both time and space and in physical terms, are of great significance for improving the ability to predict short-term algal blooms and have become a major research direction in this field. Furthermore, ensemble data assimilation algorithms, represented by Ensemble Kalman Filtering (EnKF), have become the mainstream choice for the current exploration and application of short-term algal bloom forecasting technology due to their adaptability to complex mechanistic models and their ability to fuse observational data to correct model biases.

[0003] Current short-term algal bloom forecasts mainly rely on two types of model systems: one is data-driven models (such as machine learning models). These models are trained on massive amounts of historical water quality data (such as chlorophyll a concentration and nutrient content) to achieve high-precision predictions of specific water quality indicators. However, they are essentially statistical pattern recognition based on existing data and cannot explain the physical meaning of simulation results through biochemical mechanisms. Furthermore, when future meteorological (such as sunlight and rainfall) and hydrological (such as runoff and water level) environmental conditions exceed the range of historical data, the model's generalization ability will significantly decrease. The other type is mechanism-driven models (such as three-dimensional hydrodynamic-water quality models). These models include biochemical and physical processes such as algal growth, nutrient cycling, and hydrodynamic migration. They can provide spatiotemporal continuous simulation results with physical logic and can predict the response of algae to environmental changes from a mechanistic perspective. However, the complex model structure and the large number of parameters involved lead to high uncertainty in the simulation results, which directly restricts the accuracy of short-term algal bloom forecasts. To compensate for the shortcomings of the two types of models, ensemble data assimilation techniques (such as EnKF) are gradually being promoted and applied. By introducing observational data into the mechanistic model, it attempts to integrate the advantages of both to balance forecast accuracy and interpretability.

[0004] However, current ensemble data assimilation algorithms, represented by EnKF, still face significant bottlenecks in practical applications, especially when adapting to three-dimensional hydrodynamic-water quality mechanism models. Firstly, there is a trade-off between accuracy and time cost in the model uncertainty estimation process. Because three-dimensional hydrodynamic-water quality mechanism models involve complex biochemical processes, the simulation results exhibit strong spatial heterogeneity. To avoid spurious correlations caused by small sample ensembles, traditional EnKF typically requires at least 100 ensemble members to cover the model's uncertainties. Parallel computation of a large number of ensemble members significantly increases hardware resource consumption and time. Firstly, the cost is insufficient to meet the actual needs of efficient short-term forecasting response. Secondly, the ability to capture spatial changes in algal blooms under complex environments is inadequate. Traditional EnKF mostly adopts a global assimilation strategy, without adapting and optimizing for spatial characteristics such as nutrient status and water quality background values ​​of the target water body. In complex water bodies such as eutrophic lakes, it is impossible to accurately distinguish the differences in algal bloom evolution in different areas (such as nearshore and lake center, shallow and deep water areas), resulting in forecast results that are difficult to accurately reflect the spatial distribution characteristics of algal blooms. In particular, in the case of local algal bloom outbreaks, forecast bias is prone to occur, which restricts the accuracy of short-term algal bloom prediction. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a short-term water quality forecasting method based on parameter perturbation-localized ensemble data assimilation, which can improve the accuracy of short-term water quality forecasts, save computational costs, accurately capture the spatiotemporal heterogeneity of algal blooms, and has strong interpretability and generalizability.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A short-term water quality forecasting method based on parameter perturbation-localized ensemble data assimilation includes the following steps:

[0008] S1. Based on the water quality state and its uncertainty estimation at the initial time T1 of the hydrodynamic-water quality mechanism model, construct an initial state vector, convert the initial state vector into an initial set member, and input it into the hydrodynamic-water quality mechanism model;

[0009] S2. Perform global sensitivity analysis on the water quality parameters of the hydrodynamic-water quality mechanism model to obtain a sensitive parameter set; perturb the sensitive parameter set and the model driving conditions respectively to generate a parameter perturbation set; merge the initial set members obtained in S1 with the parameter perturbation set to form a perturbation set;

[0010] S3. Input the disturbance set obtained in S2 into the hydrodynamic-water quality mechanism model, drive the model to run to the subsequent time T2, and generate the water quality forecast set at time T2.

[0011] S4. Obtain the water quality observation data and its observation error at time T2. Based on the water quality forecast set, water quality observation data and observation error, perform analysis and calculation using the ensemble Kalman filter (EnKF) algorithm to obtain the analysis set. The average value of the analysis set is the water quality optimal state estimate at time T2.

[0012] If a localization scheme is adopted, before performing the EnKF algorithm analysis and calculation, the localization function is first adjusted according to the nutrient status and water quality background value of the target water body, and then the optimal localization radius is determined by fitting the water quality observation data of the target water body. Finally, the influence weight of the observation value on different spatial locations is calculated based on the localization function and the optimal localization radius, and the influence weight is integrated into the EnKF analysis and calculation process.

[0013] S5. Update the initial conditions of the hydrodynamic-water quality mechanism model at time T2 using the analysis set obtained in S4. Repeat S2 to S4 and run the model sequentially to subsequent times T3, T4...Tn, continuously outputting the optimal water quality state estimate at each time to achieve short-term water quality forecasting.

[0014] Preferably, in S1, converting the initial state vector into initial set members specifically includes:

[0015] Obtain historical water quality simulation data of the hydrodynamic-water quality mechanism model, and generate an initial state vector matrix based on the historical water quality simulation data;

[0016] Perform empirical orthogonal function decomposition on the initial state vector matrix and select the dominant empirical orthogonal function components;

[0017] The dominant empirical orthogonal function components are converted into initial set members through second-order precise sampling.

[0018] Preferably, in S1, the number of members of the initial set is determined based on the number of computer cores and the running time cost, and the number of members of the initial set is 48.

[0019] Preferably, the hydrodynamic-water quality mechanism model is a one-dimensional, two-dimensional, or three-dimensional hydrodynamic-water quality mechanism model.

[0020] Preferably, in S1, the water quality state at the initial time T1 includes one or more of the following: chlorophyll a concentration, total phosphorus concentration, total nitrogen concentration, ammonia nitrogen concentration, and dissolved oxygen concentration.

[0021] Preferably, in S2, a global sensitivity analysis is performed on the water quality parameters of the hydrodynamic-water quality mechanism model, and the sensitive parameter set obtained specifically includes:

[0022] The Sobol sensitivity analysis method was used to calculate the global sensitivity index (ST) of each water quality parameter to the core water quality indicators in the hydrodynamic-water quality mechanism model. The ST values ​​of each water quality parameter were statistically analyzed, and parameters whose cumulative ST contribution accounted for no less than 90% of the total ST of all water quality parameters were selected to form the sensitive parameter set. The core water quality indicators include one or more of chlorophyll a concentration, dissolved oxygen concentration, total nitrogen concentration, total phosphorus concentration, and ammonia nitrogen concentration.

[0023] Preferably, in S2, perturbing the sensitive parameter set and the model driving conditions specifically includes:

[0024] The sensitive parameter set is sampled and perturbed using the Latin hypercube sampling method. The sampling perturbation follows a log-normal distribution, and the relative standard deviation of the log-normal distribution is determined based on the water quality forecast accuracy, ensemble dispersion, and ensemble quality. The model driving conditions include one or more of meteorological and hydrological data, and the perturbation of the model driving conditions is a random perturbation based on their actual fluctuation range.

[0025] Preferably, in S4, adjusting the localization function based on the trophic status and background water quality of the target water body specifically includes:

[0026] The localization function adopts the Gaspari-Cohn fifth-order polynomial correlation function; water quality observation data of different spatial points of the target water body are collected, the correlation coefficient of the observation data and its variation characteristics with spatial distance are calculated, and the baseline stable value of the correlation coefficient is determined; based on the baseline stable value and variation characteristics, the Gaspari-Cohn fifth-order polynomial correlation function is scaled and offset to obtain a localization function adapted to the target water body.

[0027] Preferably, in S4, determining the optimal localization radius by fitting the water quality observation data of the target water body specifically includes:

[0028] Historical water quality observation data from multiple observation stations of the target water body are obtained, and the correlation coefficient of water quality parameters between different observation stations and the spatial distance between the stations are calculated. The least squares method is used to fit the correlation with the adjusted localization function, and the localization radius that minimizes the fitting error is obtained as the optimal localization radius.

[0029] Preferably, in S4, the analysis and calculation performed using the ensemble Kalman filter (EnKF) algorithm based on the water quality forecast set, water quality observation data, and observation errors specifically includes:

[0030] Based on the aforementioned water quality forecast set, calculate the prior mean and prior covariance matrix of the model state;

[0031] Obtain the observation error of water quality observation data and construct the observation error covariance matrix;

[0032] Based on the prior mean, prior covariance matrix and observation error covariance matrix, the state update equation of the EnKF algorithm is constructed.

[0033] Solving the state update equation yields the posterior mean and posterior covariance matrix of the analysis set. The posterior mean is the average value of the analysis set, corresponding to the optimal water quality state estimate at time T2.

[0034] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0035] (1) This invention uses a multi-parameter perturbation strategy to perturb the sensitive parameter set (screened by global sensitivity analysis) and model driving conditions of the hydrodynamic-water quality mechanism model. Combined with the optimized design that only requires 48 set members, it solves the problem that the traditional set data assimilation requires at least 100 set members to avoid false correlations in small samples, which leads to high time costs. At the same time, it significantly improves the short-term forecast accuracy of core water quality indicators such as chlorophyll a, and achieves a balance between forecast accuracy and computational efficiency.

[0036] (2) This invention uses a localization scheme to adjust the localization function based on the nutrient status and water quality background value of the target water body, combines the least squares method to fit the optimal localization radius, and incorporates the influence weight of the observation value into the analysis and calculation of the ensemble Kalman filter (EnKF) algorithm. This solves the problem that traditional ensemble data assimilation is difficult to accurately capture the spatiotemporal changes of algal blooms in complex environments, and realizes the accurate capture of the spatiotemporal heterogeneity of algal blooms. It is especially suitable for water body scenarios with strong spatiotemporal heterogeneity.

[0037] (3) Based on the hydrodynamic-water quality mechanism model framework, this invention can trace the cause of algal blooms and identify the dominant mechanism process through statistical analysis of the simulation set. At the same time, with flexible input and output interfaces, it can access various environmental driving data and adapt to various mechanism models, solving the problems of weak interpretability and limited generalization of some forecasting methods. It achieves the unity of strong interpretability and high generalization, providing technical support for the refined management of water quality in lakes, reservoirs and other water bodies. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1This is a flowchart of a short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to the present invention;

[0040] Figure 2 A comparison chart of the prediction effects of different assimilation experiments (Assim1-4) and mechanistic models (FreeRun) provided in Embodiment 1 of the present invention;

[0041] Figure 3 This is a comparison chart of the Chl-a assimilation results (Assim1-4) of different assimilation experiments provided in Embodiment 1 of the present invention with the observed values. Detailed Implementation

[0042] 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.

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] like Figure 1 As shown, this invention provides a short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation, comprising the following steps:

[0045] S1. Based on the water quality state and its uncertainty estimation at the initial time T1 of the hydrodynamic-water quality mechanism model, construct an initial state vector, convert the initial state vector into an initial set member, and input it into the hydrodynamic-water quality mechanism model.

[0046] Specifically, converting the initial state vector into initial set members includes:

[0047] Obtain historical water quality simulation data of the hydrodynamic-water quality mechanism model, and generate an initial state vector matrix based on the historical water quality simulation data;

[0048] Perform empirical orthogonal function decomposition on the initial state vector matrix and select the dominant empirical orthogonal function components;

[0049] The dominant empirical orthogonal function components are converted into initial set members through second-order precise sampling.

[0050] The number of members in the generated initial set is determined based on factors such as the number of computer cores and runtime costs, and is generally around 50. The hydrodynamic-water quality mechanism model can be one-dimensional, two-dimensional, or three-dimensional. The water quality state at the initial time T1 includes one or more of the following: chlorophyll a concentration, total phosphorus concentration, total nitrogen concentration, ammonia nitrogen concentration, and dissolved oxygen concentration.

[0051] S2. Perform global sensitivity analysis on the water quality parameters of the hydrodynamic-water quality mechanism model to obtain a set of sensitive parameters; perturb the set of sensitive parameters and the model driving conditions respectively to generate a set of parameter perturbations; merge the initial set members obtained in S1 with the set of parameter perturbations to form a perturbation set.

[0052] Specifically, a global sensitivity analysis is performed on the water quality parameters of the hydrodynamic-water quality mechanism model, and the resulting sensitive parameter set includes:

[0053] The Sobol sensitivity analysis method was used to calculate the global sensitivity index (ST) of each water quality parameter to the core water quality indicators in the hydrodynamic-water quality mechanism model. The ST values ​​of each water quality parameter were statistically analyzed, and parameters whose cumulative ST contribution accounted for no less than 90% of the total ST of all water quality parameters were selected to form the sensitive parameter set. The core water quality indicators include one or more of chlorophyll a concentration, dissolved oxygen concentration, total nitrogen concentration, total phosphorus concentration, and ammonia nitrogen concentration.

[0054] The perturbation of the sensitive parameter set and model driving conditions specifically includes:

[0055] The sensitive parameter set is sampled and perturbed using the Latin hypercube sampling method. The sampling perturbation follows a log-normal distribution, and the relative standard deviation of the log-normal distribution is determined based on the water quality forecast accuracy, ensemble dispersion, and ensemble quality. The model driving conditions include one or more of meteorological and hydrological data, and the perturbation of the model driving conditions is a random perturbation based on their actual fluctuation range.

[0056] S3. Input the disturbance set obtained in S2 into the hydrodynamic-water quality mechanism model, drive the model to run to the subsequent time T2, and generate the water quality forecast set at time T2.

[0057] S4. Obtain the water quality observation data and its observation error at time T2. Based on the water quality forecast set, water quality observation data and observation error, perform analysis and calculation using the ensemble Kalman filter (EnKF) algorithm to obtain the analysis set. The average value of the analysis set is the water quality optimal state estimate at time T2.

[0058] If a localization scheme is adopted, before performing the EnKF algorithm analysis and calculation, the localization function is first adjusted according to the nutrient status and water quality background value of the target water body, and then the optimal localization radius is determined by fitting the water quality observation data of the target water body. Finally, the influence weight of the observation value on different spatial locations is calculated based on the localization function and the optimal localization radius, and the influence weight is integrated into the EnKF analysis and calculation process.

[0059] Specifically, adjusting the localization function based on the trophic status and background water quality of the target water body includes:

[0060] The localization function adopts the Gaspari-Cohn fifth-order polynomial correlation function; water quality observation data of different spatial points of the target water body are collected, the correlation coefficient of the observation data and its variation characteristics with spatial distance are calculated, and the baseline stable value of the correlation coefficient is determined; based on the baseline stable value and variation characteristics, the Gaspari-Cohn fifth-order polynomial correlation function is scaled and offset to obtain a localization function adapted to the target water body.

[0061] Determining the optimal localization radius by fitting water quality observation data of the target water body specifically includes:

[0062] Historical water quality observation data from multiple observation stations of the target water body are obtained, and the correlation coefficient of water quality parameters between different observation stations and the spatial distance between the stations are calculated. The least squares method is used to fit the correlation with the adjusted localization function, and the localization radius that minimizes the fitting error is obtained as the optimal localization radius.

[0063] More specifically, based on the aforementioned water quality forecast set, water quality observation data, and observation errors, the analysis and calculation performed using the ensemble Kalman filter (EnKF) algorithm specifically includes:

[0064] Based on the aforementioned water quality forecast set, calculate the prior mean and prior covariance matrix of the model state;

[0065] Obtain the observation error of water quality observation data and construct the observation error covariance matrix;

[0066] Based on the prior mean, prior covariance matrix and observation error covariance matrix, the state update equation of the EnKF algorithm is constructed.

[0067] Solving the state update equation yields the posterior mean and posterior covariance matrix of the analysis set. The posterior mean is the average value of the analysis set, corresponding to the optimal water quality state estimate at time T2.

[0068] S5. Update the initial conditions of the hydrodynamic-water quality mechanism model at time T2 using the analysis set obtained in S4. Repeat S2 to S4 and run the model sequentially to subsequent times T3, T4...Tn, continuously outputting the optimal water quality state estimate at each time to achieve short-term water quality forecasting.

[0069] The above content will be further illustrated by a specific implementation method below. The described embodiments are only some embodiments of the present invention.

[0070] Example 1

[0071] This embodiment applies the aforementioned short-term water quality forecasting method based on parameter perturbation-localized ensemble data assimilation to the short-term forecasting of chlorophyll a (Chl-a) in a eutrophic lake to verify its effect on improving the accuracy of Chl-a forecasts with small ensemble sizes. The specific implementation process is as follows:

[0072] First, an initial set is generated. Based on a three-dimensional hydrodynamic-water quality mechanism model of a eutrophic lake, the daily simulation results of Chl-a, total phosphorus (TP), total nitrogen (TN), ammonia nitrogen (NH3N), and dissolved oxygen (DO) for July-September 2018-2020 (276 days) are obtained. The daily simulation results are used as one state vector to construct an initial state vector matrix containing 276 state vectors. Empirical orthogonal function decomposition is performed on this matrix to select the top 47 empirical orthogonal function components with the largest cumulative contribution. Then, the dominant components are converted into an initial set (with 48 members) through second-order precise sampling, and the initial set is input into the three-dimensional hydrodynamic-water quality mechanism model.

[0073] After the initial set is generated, the perturbation set generation stage begins. First, the Sobol sensitivity analysis method is used to evaluate the global sensitivity index (ST) of 23 water quality parameters to Chl-a, DO, TN, TP, and NH3N in a three-dimensional hydrodynamic-water quality mechanism model of a eutrophic lake. After statistically analyzing the ST values ​​of each water quality parameter, parameters whose cumulative ST contribution accounts for more than 90% of the total ST of all parameters are selected to form a sensitive parameter set. Next, the Latin hypercube sampling method is used to sample and perturb the sensitive parameter set, following a log-normal distribution with a relative standard deviation of 0.4 (this relative standard deviation is determined based on forecast accuracy, ensemble dispersion, and ensemble quality). This generates a parameter perturbation set. Simultaneously, random perturbations based on actual fluctuation ranges are applied to the model driving conditions (such as meteorological and hydrological data like air temperature, precipitation, and inflow). Finally, the initial set and the parameter perturbation set are merged to form the perturbation set used for model input.

[0074] After the perturbation set is constructed, the mechanism model forecast is carried out. The fused perturbation set is input into the three-dimensional hydrodynamic-water quality mechanism model of a eutrophic lake, driving the model to run from the initial time T1 to the subsequent time T2, generating the Chl-a water quality forecast set at time T2.

[0075] Subsequently, optimal state estimation was performed. First, Chl-a measured data at time T2 were obtained from eight water quality monitoring stations in a eutrophic lake, and the observation errors based on instrument accuracy and spatial representativeness were determined. Then, the Gaspari-Cohn fifth-order polynomial correlation function was selected as the basic localization function, and the correlation was calculated using multi-year Chl-a monthly observation data from the eight stations. The results showed that the correlation coefficient tended to stabilize as the distance between stations increased (the baseline stable value was approximately 0.7). Based on this characteristic, the basic localization function was scaled by a factor of 0.3 and then shifted upwards by 0.7 to obtain a localization function suitable for the lake. Simultaneously, the minimum... The two-square method fits the correspondence between "inter-site correlation coefficient and spatial distance" with the adapted localization function to determine the optimal localization radius as 7236 meters. Finally, based on the Chl-a concentration forecast set at time T2, the prior mean and prior covariance matrix of the model state are calculated. Combined with the Chl-a measured data and observation errors, the observation error covariance matrix is ​​constructed. The spatial influence weight of the observed values ​​calculated by the localization function is incorporated into the EnKF algorithm to construct and solve the state update equation, thereby obtaining the Chl-a analysis set at time T2. The average value of this set is the optimal state estimate of Chl-a in a certain eutrophic lake at time T2.

[0076] Next, state updates and iterations are performed. The initial conditions of the three-dimensional hydrodynamic-water quality mechanism model are updated using the Chl-a analysis set at time T2, driving the model to run to the next time T3. Then, the above steps of "perturbation set generation → mechanism model prediction → optimal state estimation" are repeated and iterated for 3 months to finally obtain the optimal estimate of daily Chl-a (i.e., assimilation result) of a certain eutrophic lake during the study period.

[0077] To verify the effectiveness of the multi-parameter perturbation and localization schemes, this embodiment designed four assimilation experiments: Assim1 (no parameter perturbation, no localization), Assim2 (perturbation only of the most sensitive parameter—maximum algal growth rate, no localization), Assim3 (perturbation of the lumped sensitive parameters, no localization), and Assim4 (perturbation of the lumped sensitive parameters + localization scheme). Figure 2Experimental results show that compared with simulations by the three-dimensional hydrodynamic-water quality mechanism model alone, all four experiments improved the real-time prediction capability of Chl-a, indicating its ability to correct model trajectories. Compared with model simulation, the root mean square error (RMSE) of these four experiments decreased by 14%-60%, and the consistency correlation coefficient (CCC) increased by 7-14 times. Among them, the Assim4 experiment performed best due to its use of multi-parameter perturbation and localization scheme, with the highest CCC (0.92), the lowest RMSE (15.08), and the standard deviation (29.38) closest to the observed value (37.19). Specifically, the RMSE of Assim2 and Assim3 decreased by 14% and 20% respectively compared with Assim1, proving that perturbations of multiple sensitive parameters can significantly improve the Chl-a prediction effect; the RMSE of Assim4 decreased by 42% compared with Assim3, demonstrating the superior effect of the localization scheme. Figure 3 As shown, Assim4 can better simulate the extremely high values ​​of chlorophyll a (during the bloom period) and the extremely low values ​​(during the non-bloom period), which means it has a superior ability to capture the spatiotemporal heterogeneity of blooms.

[0078] Therefore, the above-mentioned method for short-term water quality forecasting based on parameter perturbation-localized ensemble data assimilation can improve the accuracy of short-term water quality forecasting, save computational costs, accurately capture the spatiotemporal heterogeneity of algal blooms, and has strong interpretability and generalizability.

[0079] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A short-term water quality forecasting method based on parameter perturbation-localized ensemble data assimilation, characterized in that, Includes the following steps: S1. Based on the water quality state and its uncertainty estimation at the initial time T1 using the hydrodynamic-water quality mechanism model, construct an initial state vector, convert the initial state vector into an initial set member, and input it into the hydrodynamic-water quality mechanism model; S2. Perform global sensitivity analysis on the water quality parameters of the hydrodynamic-water quality mechanism model to obtain a sensitive parameter set; perturb the sensitive parameter set and the model driving conditions respectively to generate a parameter perturbation set; merge the initial set members obtained in S1 with the parameter perturbation set to form a perturbation set; S3. Input the disturbance set obtained in S2 into the hydrodynamic-water quality mechanism model, drive the model to run to the subsequent time T2, and generate the water quality forecast set at time T2. S4. Obtain the water quality observation data and its observation error at time T2. Based on the water quality forecast set, water quality observation data and observation error, perform analysis and calculation using the ensemble Kalman filter (EnKF) algorithm to obtain the analysis set. The average value of the analysis set is the water quality optimal state estimate at time T2. If a localization scheme is adopted, before performing the EnKF algorithm analysis and calculation, the localization function is first adjusted according to the nutrient status and water quality background value of the target water body, and then the optimal localization radius is determined by fitting the water quality observation data of the target water body. Finally, the influence weight of the observation value on different spatial locations is calculated based on the localization function and the optimal localization radius, and the influence weight is integrated into the EnKF analysis and calculation process. In S4, adjusting the localization function based on the nutrient status of the target water body and the background water quality specifically includes: The localization function adopts the Gaspari-Cohn fifth-order polynomial correlation function; water quality observation data of different spatial points of the target water body are collected, the correlation coefficient of the observation data and its variation characteristics with spatial distance are calculated, and the baseline stable value of the correlation coefficient is determined; based on the baseline stable value and variation characteristics, the Gaspari-Cohn fifth-order polynomial correlation function is scaled and offset to obtain a localization function adapted to the target water body. Determining the optimal localization radius by fitting water quality observation data of the target water body specifically includes: Historical water quality observation data from multiple observation stations of the target water body are obtained, and the correlation coefficient between water quality parameters and spatial distance between different observation stations is calculated. The least squares method is used to fit the correlation with the adjusted localization function, and the localization radius that minimizes the fitting error is obtained as the optimal localization radius. S5. Update the initial conditions of the hydrodynamic-water quality mechanism model at time T2 using the analysis set obtained in S4. Repeat S2 to S4 and run the model sequentially to subsequent times T3, T4...Tn, continuously outputting the optimal water quality state estimate at each time to achieve short-term water quality forecasting.

2. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 1, characterized in that, In S1, converting the initial state vector into initial set members specifically includes: Obtain historical water quality simulation data of the hydrodynamic-water quality mechanism model, and generate an initial state vector matrix based on the historical water quality simulation data; Perform empirical orthogonal function decomposition on the initial state vector matrix and select the dominant empirical orthogonal function components; The dominant empirical orthogonal function components are converted into initial set members through second-order precise sampling.

3. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 2, characterized in that, In S1, the number of members in the initial set is determined based on the number of computer cores and the running time cost, and the number of members in the initial set is 48.

4. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 2, characterized in that, The hydrodynamic-water quality mechanism model is one of the following: one-dimensional, two-dimensional, or three-dimensional.

5. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 1, characterized in that, In S1, the water quality status at the initial time T1 includes one or more of the following: chlorophyll a concentration, total phosphorus concentration, total nitrogen concentration, ammonia nitrogen concentration, and dissolved oxygen concentration.

6. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 1, characterized in that, In S2, a global sensitivity analysis is performed on the water quality parameters of the hydrodynamic-water quality mechanism model, and the resulting set of sensitive parameters includes: The Sobol sensitivity analysis method was used to calculate the global sensitivity index (ST) of each water quality parameter to the core water quality indicators in the hydrodynamic-water quality mechanism model. The ST values ​​of each water quality parameter were statistically analyzed, and parameters whose cumulative ST contribution accounted for no less than 90% of the total ST of all water quality parameters were selected to form the sensitive parameter set. The core water quality indicators include one or more of chlorophyll a concentration, dissolved oxygen concentration, total nitrogen concentration, total phosphorus concentration, and ammonia nitrogen concentration.

7. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 1, characterized in that, In S2, the perturbation of the sensitive parameter set and the model driving conditions specifically includes: The sensitive parameter set is sampled and perturbed using the Latin hypercube sampling method. The sampling perturbation follows a log-normal distribution, and the relative standard deviation of the log-normal distribution is determined based on the water quality forecast accuracy, ensemble dispersion, and ensemble quality. The model driving conditions include one or more of meteorological and hydrological data, and the perturbation of the model driving conditions is a random perturbation based on their actual fluctuation range.

8. The short-term water quality forecasting method based on parameter perturbation-localization ensemble data assimilation according to claim 1, characterized in that, In S4, based on the aforementioned water quality forecast set, water quality observation data, and observation errors, the ensemble Kalman filter (EnKF) algorithm is used to perform analysis and calculations, specifically including: Based on the aforementioned water quality forecast set, calculate the prior mean and prior covariance matrix of the model state; Obtain the observation error of water quality observation data and construct the observation error covariance matrix; Based on the prior mean, prior covariance matrix and observation error covariance matrix, the state update equation of the EnKF algorithm is constructed. Solving the state update equation yields the posterior mean and posterior covariance matrix of the analysis set. The posterior mean is the average value of the analysis set, corresponding to the optimal water quality state estimate at time T2.

Images