Surface water prediction system and method based on hydrological optimization Informer model

Abstract

The invention discloses a surface water prediction system and method based on a hydrological optimization Informer model. The method comprises the following steps: collecting multi-source data of an upstream monitoring point and a downstream monitoring point of surface water; preprocessing the multi-source data to obtain preprocessed data; performing feature extraction on the preprocessed data to obtain water quality features and hydrological spatial-temporal features of the upstream monitoring point and the downstream monitoring point, and constructing accumulated dissolved oxygen saturation features; inputting the upstream and downstream water quality characteristics, the hydrological spatial-temporal characteristics and the dissolved oxygen saturation characteristics into a pre-trained multi-modal prediction model of an improved Informer model to obtain a prediction result of the upstream surface water and a prediction result of the downstream surface water; the multi-modal prediction model of the improved Informer model outputs a prediction result of the surface water obtained through fusion; according to the method, the Informer model is optimized, the GRU model is combined, hydrological spatial-temporal characteristics and accumulated dissolved oxygen saturation are introduced by utilizing an attention mechanism, the prediction precision is improved, errors are reduced, and the early warning capability for sudden pollution events is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of surface water prediction for improving the Informer model. Specifically, it relates to a surface water prediction system and method based on a hydrological optimized Informer model. Background Art

[0002] Surface water is an important part of water resources. The changes in its quantity and quality have a profound impact on the ecosystem, agricultural production, industrial water use, and human health. With the intensification of global climate change and human activities, the dynamic changes of surface water are becoming increasingly complex. Accurately predicting the changes in surface water is of great significance for the rational utilization of water resources, the protection of water environment, flood control and drought resistance, etc.

[0003] Traditional surface water prediction methods mainly rely on the assumptions of physical processes and predict the changes in surface water by simulating hydrological processes such as precipitation, evaporation, and runoff. These methods include:

[0004] Hydrological models: such as the linear ARDL model and the non - linear Informer model. Although these models can capture certain hydrological characteristics, they have limitations in dealing with complex non - linear and non - stationary data;

[0005] Time - series models: such as the ARMA model. Although it can use historical data for prediction, its computational complexity is relatively high when dealing with long - sequence data, and it is difficult to capture global characteristics;

[0006] Existing models consider hydrological conditions quite limitedly, often ignoring the continuous impact of environmental factors such as hydrological spatio - temporal characteristics and dissolved oxygen saturation in water on surface water quality, resulting in the prediction model being unable to accurately predict the changing trends of the quantity, quality, and water level of surface water. Summary of the Invention

[0007] The purpose of the present invention is to provide a surface water prediction system and method based on a hydrological optimized Informer model to solve the above - mentioned problems existing in the prior art.

[0008] Specifically, this application is as follows:

[0009] A surface water prediction method based on a hydrological optimized Informer model, comprising the following steps:

[0010] S1. Collect multi - source data related to upstream and downstream monitoring points of surface water, and the data includes hydrological data, meteorological data, terrain data, and water quality data;

[0011] S2. Pre - process the multi - source data to obtain pre - processed data; the pre - processing includes data cleaning, normalization processing, and missing value filling;

[0012] S3. Extract the water quality characteristics and hydrological spatio-temporal characteristics of the upstream and downstream monitoring points from the preprocessed data, and construct the cumulative dissolved oxygen saturation characteristics. The hydrological spatio-temporal characteristics include hydrological time characteristics and hydrological spatial characteristics;

[0013] S4. Input the water quality characteristics, hydrological spatio-temporal characteristics, and dissolved oxygen saturation characteristics of the upstream and downstream into the multi-modal prediction model of the pre-trained improved Informer model to obtain the prediction results of the upstream surface water and the prediction results of the downstream surface water. Use the mean square error method to fuse the prediction results of the upstream surface water and the prediction results of the downstream surface water, and optimize the model weights using the genetic algorithm to make the prediction results of the multi-modal prediction model of the improved Informer model adjustable in different scenarios;

[0014] S5. The multi-modal prediction model of the improved Informer model outputs the fused prediction results of the surface water. Set a preset warning threshold. If the prediction results meet the preset warning threshold, early warning of the sudden situation of the surface water at the monitoring point will be given; the prediction results include changes in water volume, water quality, and water level.

[0015] Further, step S3 includes the following steps:

[0016] Perform adversarial perturbations on all water quality characteristics and hydrological spatio-temporal characteristics in the preprocessed data to create a random forest tree; calculate the out-of-sample data normalized prediction accuracy For the out-of-sample data feature X e Apply an adversarial perturbation factor and calculate the normalized prediction accuracy after the e-th feature is perturbed Calculate X e importance P e , and the calculation expression is:

[0017]

[0018] where P e is the importance of the e-th feature, and G is the number of training samples;

[0019] Sort P e in ascending order to obtain the feature importance ranking Q e ; then construct a SelectKBest scoring system; input the sorted feature Q e into the SelectKBest scoring system to obtain the final SelectKBest scoring system score for each feature, and select features as the subsequent model input based on the SelectKBest scoring system score.

[0020] Further, the calculation method of the cumulative dissolved oxygen saturation feature in step S3 includes:

[0021] Record the reverse order of the dissolved oxygen saturation data in the cumulative period to obtain the dissolved oxygen saturation data set, calculate the influence on the dissolved oxygen saturation at time t according to Henry's law, and the calculation expression is:

[0022]

[0023] Among them, BF t is the air pressure influence intensity at time t, ABC is the saturated dissolved oxygen concentration per unit time, and T t is the air pressure intensity at time t;

[0024] Calculate the time weight at time t, and the calculation expression is:

[0025] BS t = 10 a*(t-lt) ,

[0026] Among them, BS t is the time weight, lt is the time delay (h), and a is the time weight coefficient;

[0027] Calculate the cumulative dissolved oxygen saturation excluding the time delay, and the calculation expression is:

[0028]

[0029] Among them, AR is the cumulative dissolved oxygen saturation, f represents the starting point at time t = f, and a1 represents the end point at time t = a1.

[0030] Further, the multi-modal prediction model that improves the Informer model in step S4 includes a GRU model and an Informer model. The GRU model obtains the prediction result of the upstream monitoring point according to the water quality characteristics, hydrological spatio-temporal characteristics, and cumulative dissolved oxygen saturation characteristics of the upstream monitoring point. The Informer model obtains the prediction result of the downstream surface water according to the water quality characteristics and hydrological spatio-temporal characteristics of the downstream monitoring point.

[0031] Further, the specific steps of the training process of the pre-trained multi-modal prediction model that improves the Informer model in step S4 include:

[0032] S41. Input the water quality characteristics, hydrological time characteristics in the hydrological spatio-temporal characteristics, and cumulative dissolved oxygen saturation characteristics of the upstream monitoring point into the GRU model to obtain the prediction result at the upstream monitoring point;

[0033] S42. Input the water quality characteristics, hydrological time characteristics and hydrological space characteristics in the hydrological spatio-temporal characteristics of the downstream monitoring point into the Informer model to obtain the prediction result of the downstream surface water;

[0034] S43. Take the Informer model and the GRU model as the basic training models, and perform model fusion according to the mean square error method to obtain a multi-modal prediction model for the improved Informer model.

[0035] Further, in step S41, the GRU model is an improved GRU model, and the improvement method is specifically as follows:

[0036] S411. Convert the input feature T i into a serialized time data set L Ti ;

[0037] S412. Use the attention mechanism to improve the information acquisition of the GRU algorithm;

[0038] Obtain the hidden state obtained when the GRU processes L Ti , and the hidden state is represented by G=(g i-k ,…,g i ). Use two-dimensional convolution to extract the time format matrix, and the calculation expression is:

[0039]

[0040] where C is the iterator, is the eigenvalue extracted by the q-th iterator at the m-th moment, k is the preset iterator length, n is the padding length of the two-dimensional convolution calculation, and the time format includes a short window of 6 hours, a medium window of 12 hours, and a long window of 24 hours;

[0041] Obtain the attention formula by calculating the weight:

[0042]

[0043] where MT i is the attention feature vector, σ is the Sigmoid activation function, MA is the weight matrix, n1 represents at the n1-th moment, g i represents the hidden state corresponding to the serialized time data set L Ti ;

[0044] Add the mapped MT i and the g i matrices to obtain the final predicted value;

[0045] S413. Use the genetic algorithm combined with the whale algorithm to optimize the time step, learning rate of the GRU model, and the filter length in the attention mechanism;

[0046] S414. Input the optimized attention vector, time step, learning rate, and iterator length into the GRU model to obtain the prediction result.

[0047] Furthermore, in step S42, the Informer model is an improved Informer model, and the improvement method is specifically as follows:

[0048] S421. Construct time feature encoding. Decompose the hydrological time feature in the hydrological spatio-temporal feature through EMD to obtain discrete IMF components and continuous IMF components. Use the sine function feature encoding method to generate the feature encoding at time d for the discrete IMF components:

[0049]

[0050] where Z is the time encoding method, is the sine function encoding under the Z time encoding method, per Z is the discrete IMF component, and d represents the preset period length under the Z time encoding method;

[0051] Use the CountEncoder encoding method to generate the feature encoding at time d for the continuous IMF components;

[0052] S422. Construct the longitude and latitude encoding of the spatial feature. Use the Gaussian kernel function mapping to encode the hydrological spatial feature in the hydrological spatio-temporal feature and encode the basin topological relationship:

[0053]

[0054] where c lon , c lat respectively represent the longitude coordinate and latitude coordinate of the basin center coordinate, lon i , lat i respectively represent the longitude coordinate and latitude coordinate of the monitoring point, β represents the spatial attenuation coefficient, represents the spatial feature encoding of the x-th monitoring point, and the hydrological spatial feature includes the longitude and latitude coordinate vector of the monitoring point and the key image frame feature;

[0055]

[0056] where Y (H1+L1) represents the characterization vector after H1 graph convolutions and L1 graph poolings, and are the associated characterizations of the key image frame and its neighboring key image frame in the H1-th graph convolution respectively. Yconv(·) represents the graph convolution layer, ReLU represents the activation function, and Ypool(·) represents the graph pooling layer;

[0057] S424. Use a multi-scale three-layer perceptron MLP to process the input temporal feature encoding and spatial feature encoding, and process them into a long-window temporal feature set L, a medium-window temporal feature set M, a short-window temporal feature set S, and a spatial feature set D respectively, to obtain the final multi-scale perceptual fusion feature:

[0058]

[0059] Among them, E represents the hydrological spatial feature encoding in the hydrological spatio-temporal feature, H represents the key image frame feature encoding in the hydrological spatial feature, NOR represents the normalization algorithm, MLP represents the multi-layer perceptron operation, b1 represents different scale feature sets, denotes concatenation, G′ LMSD denotes the normalized fusion feature, G LMSD denotes the final multi-scale perceptual fusion feature. The spatial feature set D represents the final spatial embedding feature vector obtained by concatenating the hydrological spatial feature encoding E and the key image frame feature encoding H in the hydrological spatio-temporal feature through the multi-scale perceptron MLP;

[0060] S425. Input the obtained multi-scale perceptual fusion feature into the Informer model as the input feature encoding to obtain an improved Informer model. The improved Informer model also introduces a window sparse self-attention mechanism, a distillation mechanism, and a fully connected output mechanism. The output of the convolutional neural network is used to access the first input channel of the improved Informer model.

[0061] Furthermore, the window sparse self-attention mechanism in step 425 includes:

[0062] Characterize the self-attention mechanism using a sliding window form combined with the Fourier transform;

[0063] Use a sliding window form to limit the scope of action of the attention mechanism so that it only calculates the attention weights within a local window;

[0064] Use the Fourier transform to transform the time series data from the time domain to the frequency domain, which can not only capture local time fluctuations but also handle global trends.

[0065] Furthermore, the calculation expression for model fusion according to the mean square error method is:

[0066] f F =W I *f d +W T *f D ,

[0067]

[0068] Among them, f d is the predicted value of the upstream surface water obtained by the Informer model, and f D is the predicted value of the downstream surface water obtained by the GRU model, and f F is the final predicted value after fusion. W I and W T are the weight coefficients of the Informer model and the GRU model respectively. θ I and θ L represent the mean square error of the prediction results of the Informer model and the GRU model respectively.

[0069] A surface water prediction system based on a hydrological optimized Informer model. The system is used to execute any one of the surface water prediction methods based on a hydrological optimized Informer model. The system includes a data collection module, a data preprocessing module, a feature extraction module, a model training module, and a model prediction module;

[0070] The data collection module is used to collect multi-source data related to upstream and downstream monitoring points of surface water. The data includes hydrological data, meteorological data, topographic data, and water quality data;

[0071] The data preprocessing module is used to preprocess the multi-source data to obtain preprocessed data. The preprocessing includes data cleaning, normalization processing, and missing value filling;

[0072] The feature extraction module is used to extract features from the preprocessed data to obtain water quality features and hydrological spatio-temporal features of upstream and downstream monitoring points, and construct cumulative dissolved oxygen saturation features. The hydrological spatio-temporal features include hydrological time features and hydrological space features;

[0073] The model training module is used to input the water quality features, hydrological spatio-temporal features, and dissolved oxygen saturation features of the upstream and downstream into the multi-modal prediction model of the pre-trained improved Informer model to obtain the prediction results of the upstream surface water and the prediction results of the downstream surface water, and fuse the prediction results of the upstream surface water and the downstream surface water using the mean square error method;

[0074] The model prediction module is used to output the prediction results of the surface water obtained by fusing the multi-modal prediction model of the improved Informer model. The prediction results include changes in water volume, water quality, and water level.

[0075] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0076] An embodiment of the present invention provides a method for collecting multi-source data related to upstream and downstream monitoring points of surface water; preprocessing the multi-source data to obtain preprocessed data; extracting features from the preprocessed data to obtain water quality features and hydrological spatio-temporal features of the upstream and downstream monitoring points, and constructing a cumulative dissolved oxygen saturation feature; inputting the water quality features, hydrological spatio-temporal features, and dissolved oxygen saturation features of the upstream and downstream into a multi-modal prediction model of an improved Informer model that has been pre-trained to obtain prediction results of upstream surface water and prediction results of downstream surface water; and outputting and fusing the prediction results of the surface water obtained by the multi-modal prediction model of the improved Informer model. The present invention optimizes the Informer model through hydrology and introduces an attention connection of cumulative dissolved oxygen saturation and hydrological spatio-temporal features by combining with the GRU model, improves the calculation accuracy, reduces the prediction error of surface water, and gives an early warning for sudden pollution. Description of the Drawings

[0077] Figure 1 FIG. is a schematic flowchart of a method for predicting surface water based on a hydrologically optimized Informer model provided by an embodiment of the present invention;

[0078] Figure 2 FIG. is an architecture diagram of a surface water prediction system based on a hydrologically optimized Informer model provided by an embodiment of the present invention;

[0079] Figure 3 FIG. is an architecture diagram of the Informer model of a surface water prediction system based on a hydrologically optimized Informer model provided by an embodiment of the present invention. Detailed Embodiments

[0080] The present invention will be described in detail below with reference to the accompanying drawings.

[0081] Embodiment 1

[0082] First, technical terms related to the embodiments of the present application will be explained.

[0083] (1) Informer model

[0084] The Informer model is an efficient time series prediction model, especially suitable for processing long sequence data. It significantly improves the computational efficiency and prediction performance of the model by introducing a sparse self-attention mechanism and a series of optimization techniques. The Informer model performs well in fields such as meteorological prediction, financial time series analysis, and hydrological prediction.

[0085] (2) GRU model

[0086] GRU (Gated Recurrent Unit) is an improved recurrent neural network (RNN) architecture used to handle long short-term dependencies in sequential data. GRU controls the flow of information by introducing an update gate and a reset gate, thus effectively alleviating the vanishing gradient problem of traditional RNNs when dealing with long sequences. The GRU model has been widely applied in fields such as natural language processing, time series prediction, and speech recognition.

[0087] An embodiment of the present invention provides a surface water prediction method based on a hydrological optimization Informer model, as Figure 1 , including the following steps:

[0088] S1. Collect multi-source data related to upstream and downstream monitoring points of surface water, and the data includes hydrological data, meteorological data, terrain data, and water quality data;

[0089] S2. Preprocess the multi-source data to obtain preprocessed data; the preprocessing includes data cleaning, normalization processing, and missing value filling;

[0090] S3. Extract features from the preprocessed data to obtain water quality features and hydrological spatio-temporal features of upstream and downstream monitoring points, and construct cumulative dissolved oxygen saturation features. The hydrological spatio-temporal features include hydrological time features and hydrological space features;

[0091] S4. Input the water quality features, hydrological spatio-temporal features, and dissolved oxygen saturation features of upstream and downstream into the multi-modal prediction model of the pre-trained improved Informer model to obtain the prediction results of upstream surface water and downstream surface water. Use the mean square error method to fuse the prediction results of upstream surface water and downstream surface water, and adopt a genetic algorithm to optimize the model weights to make the prediction results of the multi-modal prediction model of the improved Informer model adjustable in different scenarios;

[0092] S5. The multi-modal prediction model of the improved Informer model outputs the fused prediction results of surface water. Preset an early warning threshold. If the prediction results meet the preset early warning threshold, early warning of sudden surface water conditions at the monitoring points is given in advance; the prediction results include changes in water volume, water quality, and water level.

[0093] Specifically, this method collects multi-source data related to upstream and downstream monitoring points of surface water; preprocesses the multi-source data to obtain preprocessed data; extracts features from the preprocessed data to obtain water quality features and hydrological spatio-temporal features of the upstream and downstream monitoring points, and constructs cumulative dissolved oxygen saturation features; inputs the water quality features, hydrological spatio-temporal features, and dissolved oxygen saturation features of the upstream and downstream into a multi-modal prediction model of an improved Informer model that has been pre-trained to obtain prediction results of upstream surface water and prediction results of downstream surface water; the multi-modal prediction model of the improved Informer model outputs the fused prediction results of surface water; in this embodiment, by hydrologically optimizing the Informer model and introducing attention connections of cumulative dissolved oxygen saturation and hydrological spatio-temporal features in combination with the GRU model, the calculation accuracy is improved, the prediction error of surface water is reduced, and early warnings are given for sudden pollution. The sudden conditions of surface water at least include sudden pollution of surface water, sudden floods of surface water, etc.

[0094] It should be noted that the Informer model is suitable for hydrological prediction. For example, its long sequence capture ability is better than that of other time series capture models such as LSTM. Therefore, the present invention uses the Informer model for hydrological prediction.

[0095] By introducing cumulative dissolved oxygen saturation, based on the historical time series relationship, the excellent performance of the Informer model in long sequence capture ability enhances the long-term prediction ability of the model and improves the model prediction accuracy. This is one of the technical highlights of the present invention;

[0096] It should be noted that the collection of surface water and preprocessing in steps S1 and S2 include:

[0097] Collection of original water quality time series data of each monitoring point (water quality parameters such as pH, DO, NH3-N, TP, etc., sampling frequency of each monitoring point is once every half hour): time series concentration data of each monitoring point within 24 hours;

[0098] The data collection and processing process is as follows:

[0099] Calculation of the real-time volatility of each monitoring point, calculating the hourly change rate for each collected water quality parameter Dynamically evaluate the stationarity of data collection, where r t : the real-time volatility at time t (dimensionless), x t : the value of the water quality parameter at time t (such as DO concentration, unit mg / L), x t-1 : the value of the water quality parameter at time t - 1;

[0100] Window adaptive adjustment of the collection method:

[0101] Collection during stable periods (|r t∣< 5%): Adopt a long window (24 hours), enhance the periodic characteristics, applicable to the stable period (such as the slow change of DO in the dry season), and smooth the diurnal cycle noise;

[0102] Collection during the fluctuating period (5% ≤ ∣r t ∣≤ 15%): Switch to a medium window (12 hours), balance the trend and details, cope with the initial stage of rainfall or tidal influence, and retain the semi-diurnal cycle characteristics;

[0103] Collection during the mutation period (∣r t ∣> 15%): Enable a short window (6 hours), quickly respond to sudden changes, target sudden pollution (such as factory leakage), and minimize the lag effect;

[0104] For the window adaptive adjustment of the collection method, adopt multi-scale feature generation through parallel computing:

[0105] Independently calculate the moving average values of three windows of 6h / 12h / 24h for each collected data to generate three types of time series features:

[0106]

[0107] Among them, win represents the window hours, Win ∈ {6, 12, 24};

[0108] Data preprocessing, through missing value filling, adopt mirror padding (Reflection Padding) to avoid sequence truncation;

[0109] Example of mirror padding in this embodiment:

[0110] Original DO sequence: [8.2, 8.5, 7.9,..., 9.1],

[0111] MA of 6h window (capturing hourly fluctuations): [8.3, 8.2,..., 8.9],

[0112] MA of 24h window (highlighting the daily cycle trend): [8.4, 8.4,..., 8.6].

[0113] In the above embodiment, specifically, step S3 includes the following steps:

[0114] Perform adversarial perturbation on all water quality features and hydrological spatio-temporal features in the preprocessed data, create a random forest tree; calculate the normalized prediction accuracy of out-of-sample data For the out-of-sample data feature X e Apply an adversarial perturbation factor and calculate the normalized prediction accuracy after perturbing the e-th feature Calculate the importance P of X e of e, the calculation expression is:

[0115]

[0116] where P e is the importance of the e-th feature, and G is the number of training samples;

[0117] Sort P e in ascending order to obtain the sorted feature importance Q e ; then construct the SelectKBest scoring system; input the sorted feature Q e into the SelectKBest scoring system to obtain the final SelectKBest score for each feature, and select features as the input of the subsequent model according to the SelectKBest scores.

[0118] It should be noted that by performing adversarial perturbations on water quality features and hydro-spatial-temporal features, constructing the SelectKBest scoring system, and screening features according to the scores, compared with the prior art, the training amount of the subsequent model data is reduced, the training time of the subsequent model is improved, and at the same time, the accuracy of feature screening is improved. In addition to applying adversarial perturbation factors, other algorithm factors can also be added, including but not limited to Gaussian noise factors, Poisson noise factors, etc., which will not be elaborated in this embodiment.

[0119] In the above embodiment, specifically, the calculation method of the cumulative dissolved oxygen saturation feature in step S3 includes:

[0120] Record the reverse order of the dissolved oxygen saturation data in the cumulative period to obtain the dissolved oxygen saturation data set, and calculate the influence on the dissolved oxygen saturation at time t according to Henry's law. The calculation expression is:

[0121]

[0122] where BF t is the air pressure influence intensity at time t, ABC is the saturated dissolved oxygen concentration per unit time, and T r is the air pressure intensity at time t;

[0123] Calculate the time weight at time t. The calculation expression is:

[0124] BS t = 10 a*(t-lt) ,

[0125] where BS t is the time weight, lt is the time delay, and a is the time weight coefficient;

[0126] Calculate the cumulative dissolved oxygen saturation excluding the time delay. The calculation expression is:

[0127]

[0128] Among them, AR is the cumulative dissolved oxygen saturation, f represents the starting point at time t = f, and a1 represents the ending point at time t = a1.

[0129] It should be noted that the existing models consider hydrology quite limitedly, often ignoring the continuous impact of environmental factors such as dissolved oxygen saturation on water quality disasters and sudden pollution, resulting in the prediction model being unable to accurately describe the changing trend of water quality. In this embodiment, by introducing the principle of high correlation between the decomposition of substances in water bodies and the characteristics of dissolved oxygen saturation, the characteristics of dissolved oxygen saturation are continuously analyzed through the Informer neural network, and then sudden water quality pollution accidents of surface water are predicted with high precision.

[0130] In the above embodiment, specifically, the multi-modal prediction model for improving the Informer model in step S4 includes a GRU model and an Informer model. The GRU model obtains the prediction result of the upstream monitoring point based on the water quality characteristics, hydrological spatio-temporal characteristics, and cumulative dissolved oxygen saturation characteristics of the upstream monitoring point. The Informer model obtains the prediction result of the downstream surface water based on the water quality characteristics and hydrological spatio-temporal characteristics of the downstream monitoring point.

[0131] It should be noted that in this embodiment, the GRU model and the Informer model are selected to fuse and predict surface water because the advantages of the GRU model in processing time series and the Informer model in processing spatio-temporal series are combined. Through the fusion of the two models, the accuracy of the prediction analysis of the fusion model is improved. Compared with the prior art, the fusion of these two models is the technical highlight of the present invention.

[0132] In the above embodiment, specifically, the specific steps of the training process of the pre-trained multi-modal prediction model for improving the Informer model in step S4 include:

[0133] S41. Input the water quality characteristics, hydrological time characteristics in the hydrological spatio-temporal characteristics, and cumulative dissolved oxygen saturation characteristics of the upstream monitoring point into the GRU model to obtain the prediction result of the upstream monitoring point at that time;

[0134] S42. Input the water quality characteristics, hydrological time characteristics and hydrological space characteristics in the hydrological spatio-temporal characteristics of the downstream monitoring point into the Informer model to obtain the prediction result of the downstream surface water;

[0135] S43. Take the Informer model and the GRU model as the basic training models, and perform model fusion according to the mean square error method to obtain the multi-modal prediction model for improving the Informer model.

[0136] In the above embodiments, specifically, in step S41, the GRU model is an improved GRU model, and the improvement method is specifically as follows:

[0137] S411. Convert the input feature T i into a serialized time dataset L Ti ;

[0138] S412. Use the attention mechanism to improve the information acquisition of the GRU algorithm;

[0139] Obtain the hidden state obtained when the GRU processes L Ti . The hidden state is represented by G=(g i-k ,…,g i ). Use two-dimensional convolution to extract the time format matrix, and the calculation expression is:

[0140]

[0141] where C is an iterator, is the eigenvalue extracted by the q-th iterator at the m-th moment, k is the preset iterator length, n is the padding length of the two-dimensional convolution calculation, and the time format includes a short window of 6 hours, a medium window of 12 hours, and a long window of 24 hours;

[0142] Obtain the attention formula by calculating the weight:

[0143]

[0144] where MT i is the attention feature vector, σ is the Sigmoid activation function, MA is the weight matrix, n1 represents the n1-th moment, and g i represents the serialized time dataset corresponding hidden state;

[0145] Add the mapped MT i and the g i matrices to obtain the final predicted value;

[0146] S413. Use the genetic algorithm combined with the whale algorithm to optimize the time step, learning rate, and filter length in the attention mechanism of the GRU model;

[0147] S414. Input the optimized attention vector, time step, learning rate, and iterator length into the GRU model to obtain the prediction result.

[0148] Specifically, the specific process of using the genetic algorithm to optimize the time step, learning rate, and filter length in the attention mechanism of the GRU model is as follows:

[0149] M1: Parameter encoding. The time step, learning rate, and filter length in the attention mechanism are optimized using a genetic algorithm. The time step, learning rate, and filter length in the attention mechanism are converted into binary digital form and encoded as antibodies in the genetic algorithm.

[0150] M2: Calculate the fitness function. The key issue in optimizing the time step, learning rate, and filter length in the attention mechanism is the selection of the fitness function. The integral of the absolute error is used as the evaluation index, and the fitness function is:

[0151]

[0152] M3: Update memory cells. Antibody cells with high fitness are retained through the memory function and assigned to the time step, learning rate, and filter length in the attention mechanism. Calculate the individual fitness function value and determine the optimal fitness value of the population. If the optimal fitness value searched is less than the optimal fitness value in the immune network, the immune memory starts to search for the optimal value again and serves as the optimal immune antibody. Otherwise, add the already calculated optimal fitness value to the antibody memory list.

[0153] M4: Maintain diversity. The diversity of the population is adjusted using the selection probability, and its calculation formula is

[0154]

[0155] where a and b are random numbers in the interval [0, 1], F i (i) is the individual fitness value, and C is the concentration of the antibody.

[0156] M5: Crossover operation. The individuals in the population are operated using the two-point crossover method, and its crossover equation is:

[0157]

[0158] where X′ and Y′ are the new individuals after crossover, and r is a random number in the interval [0, 1].

[0159] M6: Mutation operation. Gaussian mutation is used to operate on the time step, learning rate, and filter length in the attention mechanism, and its mutation equation is

[0160]

[0161] where λ is a random number in the interval [0, 1], and μ is the Gaussian operator.

[0162] M7: The algorithm terminates. Compare the fitness function values of two adjacent times. If the error meets the condition or the maximum number of iterations is reached, the algorithm terminates, and the obtained result is used to adjust the parameters of the time step, learning rate, and filter length in the attention mechanism in real time. Otherwise, go to step M2.

[0163] The whale algorithm in this embodiment will not be elaborated here and can be understood by referring to the existing technology.

[0164] It should be noted that by using the genetic algorithm combined with the whale algorithm to optimize the time step, learning rate, and filter length in the attention mechanism of the GRU model, the prediction accuracy of the GRU model is improved, and the prediction speed of the GRU model is accelerated.

[0165] In the above embodiment, specifically, in step S42, the Informer model is an improved Informer model, and the improvement method is specifically as follows:

[0166] S421. Construct time feature encoding. Decompose the hydrological time feature in the hydrological spatio-temporal feature through EMD to obtain discrete IMF components and continuous IMF components. Use the sine function feature encoding method to generate the feature encoding at time d for the discrete IMF components:

[0167]

[0168] Among them, Z is the time encoding method, is the sine function encoding under the Z time encoding method, per Z is the discrete IMF component, and d represents the preset period length under the Z time encoding method;

[0169] Use the CountEncoder encoding method to generate the feature encoding at time d for the continuous IMF components;

[0170] It should be noted that the sine function feature encoding and the CountEncoder encoding are respectively used for the IMF components because the CountEncoder encoding has advantages in processing continuous IMF components, while the sine function feature encoding method has advantages in encoding discrete IMF components. This can improve the encoding efficiency and accuracy and provide support for subsequent model training and recognition;

[0171] S422. Construct the longitude and latitude encoding of the spatial feature. Use the Gaussian kernel function mapping to encode the hydrological spatial feature in the hydrological spatio-temporal feature and encode the basin topological relationship:

[0172]

[0173] Among them, among them, c lon , c latrespectively represent the longitude coordinate and latitude coordinate of the basin center coordinates, lon i , lat i respectively represent the longitude coordinate and latitude coordinate of the monitoring point, β represents the spatial attenuation coefficient, represents the spatial feature encoding of the x-th monitoring point, and the hydrological spatial features include the longitude and latitude coordinate vectors of the monitoring point and the key image frame features;

[0174] It should be noted that the longitude and latitude encoding for constructing spatial features corresponds to the longitude and latitude position information of each monitoring point;

[0175]

[0176] Among them, Y (H1+L1) represents the representation vector after H1 graph convolutions and L1 graph poolings, and are respectively the associated representations of the key image frame and its neighboring key image frames in the H1-th graph convolution. Yconv(·) represents the graph convolution layer, ReLU represents the activation function, and Ypool(·) represents the graph pooling layer. The key image frame features represent the dirty image features of the surface water at the monitoring point, the water level features of the surface water at the monitoring point, and the water flow features of the surface water at the monitoring point, as Figure 2 shown;

[0177] It should be noted that the key image frame feature encoding for constructing spatial features is encoded for the key image frames collected for each monitoring point. The spatial feature encoding includes the longitude and latitude encoding and key image frame feature encoding of the spatial features corresponding to each monitoring point. The convolutional neural network described in this embodiment adopts a Feature Pyramid Network (FPN). The Feature Pyramid Network constructs a multi-scale feature pyramid and uses feature maps at different levels to process targets at different scales, thereby improving the model's processing ability for multi-scale data and improving the recognition of the dirty image features of the surface water associated with the key image frame features of the monitoring point;

[0178] Specifically, the Feature Pyramid Network in this embodiment is divided into four layers, and the specific processing process of the four layers of the Feature Pyramid Network is as follows:

[0179] The first layer of the pyramid is used to process the resolution of the first key frame image, and the brightness of the image belongs to the first-level brightness;

[0180] The second layer of the pyramid is used to process the resolution of the second key frame image, and the brightness of the image belongs to the second-level brightness;

[0181] The third layer of the pyramid is used to process the resolution of the third key frame image, and the brightness of the image belongs to the third-level brightness;

[0182] The fourth layer of the pyramid is used to process the resolution of the fourth key-frame image, and the brightness of the image belongs to the fourth-level brightness;

[0183] The resolution levels of the said image are specifically divided as follows:

[0184] Resolution of the first key-frame image: w < 150, h < 150;

[0185] Resolution of the second key-frame image: 150 < w ≤ 320, 150 < h ≤ 350;

[0186] Resolution of the third key-frame image: 350 < w ≤ 1080, 350 < h ≤ 1080;

[0187] Resolution of the fourth key-frame image: 1080 < w, 1080 < h, where w represents the width of the image and h represents the height of the image;

[0188] The brightness of the said image adopts the method of calculating the mean brightness of the image by using a histogram, and is specifically divided as follows:

[0189] First mean brightness: 0 ≤ L ≤ 70;

[0190] Second mean brightness: 70 < L ≤ 110;

[0191] Third mean brightness: 110 < L ≤ 170;

[0192] Fourth mean brightness: 170 < L ≤ 255, where L represents the mean brightness value of the image;

[0193] Two branch networks are added behind the feature maps of each layer of the pyramid. One branch is used for classification and the other branch is used for regression. And each branch first performs 6 convolutions on the feature map to enhance the key image frames respectively. The convolution kernel size is 3×3 and the number is 64. By setting the feature recognition of four different brightness levels, the recognition degree of the turbidity of the surface water of the monitoring point by the model is improved.

[0194] S424. Use a multi-scale three-layer perceptron MLP to process the input temporal feature encoding and spatial feature encoding, and process them into a long-window temporal feature set L, a medium-window temporal feature set M, a short-window temporal feature set S, and a spatial feature set D respectively, to obtain the final multi-scale perceptual fusion feature:

[0195]

[0196] Among them, E represents the hydrological spatial feature encoding in the hydrological spatio-temporal feature, H represents the key image frame feature encoding in the hydrological spatial feature, NOR represents the normalization algorithm, MLP represents the multi-layer perceptron operation, b1 represents different scale feature sets, represents splicing, G′LMSD represents the normalized fusion feature, G LMSD represents the final multi-scale perception fusion feature. The spatial feature set D represents the final spatial embedding feature vector obtained by concatenating the hydrological spatial feature encoding E and the key image frame feature encoding H in the hydrological spatio-temporal features through the multi-scale perception machine MLP;

[0197] S425. Input the obtained multi-scale perception fusion feature into the Informer model as the input feature encoding to obtain an improved Informer model. A window sparse self-attention mechanism, a distillation mechanism, and a fully connected output mechanism are also introduced in the improved Informer model. The output of the convolutional neural network is used to access the first input channel of the improved Informer model.

[0198] In the above embodiment, specifically, the window sparse self-attention mechanism in step 425 includes:

[0199] Characterize the self-attention mechanism using a sliding window form combined with the Fourier transform;

[0200] Use a sliding window form to limit the scope of action of the attention mechanism so that it calculates the attention weights only within a local window;

[0201] Use the Fourier transform to transform the time series data from the time domain to the frequency domain, which can not only capture local time fluctuations but also handle global trends.

[0202] It should be noted that the window sparse self-attention mechanism includes:

[0203] Characterize the self-attention mechanism using a window form. The original self-attention mechanism can be expressed as:

[0204]

[0205] The self-attention of the i-th q after using the window form can be expressed as:

[0206]

[0207] where Q, K, and V represent the Query, Key, and Value matrices respectively, while q, k, and v represent the corresponding vectors, softmax is the activation function, and d represents the dimension of the key.

[0208] In the above embodiment, specifically, the calculation expression for model fusion according to the mean square error method is:

[0209] f F =W I *f d +WT *f D ,

[0210]

[0211] Among them, f d is the predicted value of the upstream surface water obtained by the Informer model, and f D is the predicted value of the downstream surface water obtained by the GRU model, and f F is the final predicted value after fusion. W I and W T are the weight coefficients of the Informer model and the GRU model respectively. θ I and θ L represent the mean square errors of the prediction results of the Informer model and the GRU model respectively.

[0212] It should be noted that the specific implementation process of optimizing the model weights using the genetic algorithm in step S4 is as follows:

[0213] The genetic algorithm (Genetic Algorithm, GA) is an optimization algorithm based on the principles of natural selection and genetics. By simulating the "selection", "crossover", and "mutation" operations in the biological evolution process, it seeks the optimal solution to the problem. The following is the specific process of using the genetic algorithm to optimize the weights of the Informer model and the GRU model:

[0214] 1. Initialize the population

[0215] Population size: Randomly generate a group of populations, and each individual represents a set of hyperparameter configurations of the model.

[0216] Hyperparameter range: Set a reasonable value range for each hyperparameter, including the learning rate, the number of network layers, the number of neurons, etc.

[0217] 2. Evaluate the fitness

[0218] Model training: Use each set of hyperparameter configurations to train the Informer and GRU models;

[0219] Performance evaluation: Evaluate the model performance on the validation set. Common evaluation metrics include the mean square error (MSE), accuracy, etc.;

[0220] 3. Selection operation

[0221] Selection strategy: Select individuals with higher fitness to enter the next generation according to the evaluation performance. Common selection methods include roulette wheel selection, tournament selection, etc.;

[0222] 4. Crossover operation

[0223] Crossover strategy: Perform crossover operations on the selected individuals, exchange some of their genes (i.e., hyperparameters) to generate new individuals;

[0224] 5. Mutation operation

[0225] Mutation strategy: Modify some genes of an individual slightly with a certain probability to increase the population diversity;

[0226] 6. Iterative optimization

[0227] Iterative process: Repeat the above selection, crossover, and mutation operations until a predetermined number of iterations is reached or the stopping criterion is met;

[0228] Stopping condition: Set the maximum number of iterations or the fitness value less than the set threshold as the stopping condition;

[0229] 7. Output the optimal solution

[0230] Optimal parameters: After multiple rounds of iteration, find a set of optimal hyperparameter combinations;

[0231] Model training: Retrain the model with the optimal hyperparameters to obtain the best performance;

[0232] Example 2

[0233] A surface water prediction system based on a hydrological optimization Informer model, the system includes a data collection module A1, a data preprocessing module A2, a feature extraction module A3, a model training module A4, and a model prediction module A5;

[0234] The data collection module A1 is used to collect multi-source data related to upstream and downstream monitoring points of surface water, and the data includes hydrological data, meteorological data, topographic data, and water quality data;

[0235] The data preprocessing module A2 is used to preprocess the multi-source data to obtain preprocessed data; the preprocessing includes data cleaning, normalization processing, and missing value filling;

[0236] The feature extraction module A3 is used to extract features from the preprocessed data to obtain water quality features and hydrological spatio-temporal features of upstream and downstream monitoring points, and construct cumulative dissolved oxygen saturation features, and the hydrological spatio-temporal features include hydrological time features and hydrological space features;

[0237] The model training module A4 is used to input the water quality characteristics, hydrological spatio-temporal characteristics, and dissolved oxygen saturation characteristics of the upstream and downstream into the multi-modal prediction model of the pre-trained improved Informer model to obtain the prediction results of the upstream surface water and the prediction results of the downstream surface water. The mean square error method is used to fuse the prediction results of the upstream surface water and the prediction results of the downstream surface water, and the genetic algorithm is used to optimize the model weights, so that the prediction results of the multi-modal prediction model of the improved Informer model are adjustable in weight under different scenarios;

[0238] The model prediction module A5 is used to output the prediction results of the surface water obtained by fusing the multi-modal prediction model of the improved Informer model, and the prediction results include changes in water volume, water quality, and water level.

[0239] It should be understood that the above embodiments are one or more embodiments of the present invention, and there are many other embodiments and their deformations based on the present invention; when ordinary technicians in this industry do not make pioneering innovations, the deformations and modifications made through the present invention all fall within the protection scope of the present invention.

Claims

1. A surface water prediction method based on a hydrological optimization Informer model, characterized in that, It includes the following steps: S1. Collect multi-source data related to upstream and downstream monitoring points of surface water, where the data includes hydrological data, meteorological data, topographical data, and water quality data; S2. Preprocess the multi-source data to obtain preprocessed data; the preprocessing includes data cleaning, normalization, and missing value filling; S3. Extract features from the preprocessed data to obtain water quality features and hydrological spatio-temporal features of upstream and downstream monitoring points, and construct cumulative dissolved oxygen saturation features. The hydrological spatio-temporal features include hydrological time features and hydrological spatial features; S4. Input the water quality features, hydrological spatio-temporal features, and dissolved oxygen saturation features of upstream and downstream into the multi-modal prediction model of the pre-trained improved Informer model to obtain the prediction results of upstream surface water and downstream surface water. Use the mean square error method to fuse the prediction results of upstream surface water and downstream surface water, and adopt the genetic algorithm to optimize the model weights to make the prediction results of the multi-modal prediction model of the improved Informer model adjustable in different scenarios; S5. The multi-modal prediction model of the improved Informer model outputs the fused prediction results of surface water. Set a preset warning threshold. If the prediction results meet the preset warning threshold, early warning of sudden situations of surface water at the monitoring points will be given; the prediction results include changes in water volume, water quality, and water level.

2. The surface water prediction method based on the hydrological optimization Informer model according to claim 1, wherein, Step S3 includes the following steps: Perform adversarial perturbations on all water quality characteristics and hydro-spatial-temporal characteristics in the preprocessed data, and create a random forest tree; calculate the normalized prediction accuracy of out-of-sample data For the out-of-sample data feature X e Apply an adversarial perturbation factor and calculate the normalized prediction accuracy after perturbing the e-th feature Calculate X e importance P e , and the calculation formula is: where P e is the importance of the e-th feature, and G is the number of training samples; Sort P e in ascending order to obtain the sorted feature importance Q e ; then construct a SelectKBest scoring system; input the sorted feature Q e into the SelectKBest scoring system to obtain the final SelectKBest score for each feature, and select features based on the SelectKBest scores as the input for the subsequent model.

3. A surface water prediction method based on a hydrological optimization Informer model according to claim 1, characterized in that, The calculation method of the cumulative dissolved oxygen saturation feature in step S3 includes: Record the reverse order of the dissolved oxygen saturation data in the cumulative period to obtain the dissolved oxygen saturation data set, and calculate the influence on the dissolved oxygen saturation at time t according to Henry's law. The calculation expression is: Among them, BF t is the air pressure influence intensity at time t, ABC is the saturated dissolved oxygen concentration per unit time, T t is the air pressure intensity at time t; Calculate the time weight at time t. The calculation expression is: BS t = 10 a*(t-lt) , Among them, BS t is the time weight, lt is the time delay, and a is the time weight coefficient; Calculate the cumulative dissolved oxygen saturation excluding time delay. The calculation expression is: Where AR is the cumulative dissolved oxygen saturation, f represents the starting point at time t = f, and a1 represents the ending point at time t = a1.

4. A surface water prediction method based on a hydrological optimized Informer model according to claim 1, characterized in that, In step S4, the multi-modal prediction model of the improved Informer model includes a GRU model and an Informer model. The GRU model obtains the prediction results of the upstream monitoring point based on the water quality features, hydrological spatio-temporal features, and cumulative dissolved oxygen saturation features of the upstream monitoring point. The Informer model obtains the prediction results of downstream surface water based on the water quality features and hydrological spatio-temporal features of the downstream monitoring point.

5. A surface water prediction method based on a hydrological optimization Informer model according to claim 4, characterized in that, The specific steps of the training process of the pre-trained multi-modal prediction model of the improved Informer model in step S4 include: S41. Input the water quality features, hydrological time features in the hydrological spatio-temporal features, and cumulative dissolved oxygen saturation features of the upstream monitoring point into the GRU model to obtain the prediction results at the upstream monitoring point; S42. Input the water quality features, hydrological time features and hydrological spatial features in the hydrological spatio-temporal features of the downstream monitoring point into the Informer model to obtain the prediction results of downstream surface water; S43. Take the Informer model and the GRU model as the basic training models, and perform model fusion according to the mean square error method to obtain a multi-modal prediction model that improves the Informer model.

6. A surface water prediction method based on a hydrological optimized Informer model according to claim 5, characterized in that, In step S41, the GRU model is an improved GRU model, and the improvement method is specifically as follows: S411. Convert the input feature T i into a serialized time data set L Ti ; S412. Use the attention mechanism to improve the information acquisition of the GRU algorithm; Obtain the hidden state obtained when the GRU processes L Ti When, the hidden state is represented by G=(g i-k ,…,g i ). Use two-dimensional convolution to extract the time format matrix, and the calculation expression is: where C is an iterator, is the eigenvalue extracted by the q-th iterator at time m, k is the preset iterator length, n is the padding length for two-dimensional convolution calculation, and the time format includes a short window of 6 hours, a medium window of 12 hours, and a long window of 24 hours; Obtain the attention formula by calculating the weight: Among them, MT i is the attention feature vector, σ is the Sigmoid activation function, MA is the weight matrix, n1 represents the n1-th moment, and g i represents the serialized time dataset L Ti corresponding hidden state; Add MT i and g i after matrix mapping and then sum them up to obtain the final predicted value; S413. Use the genetic algorithm combined with the whale algorithm to optimize the time step, learning rate, and filter length in the attention mechanism of the GRU model; S414. Input the optimized attention vector, time step, learning rate, and iterator length into the GRU model to obtain the prediction result.

7. A surface water prediction method based on a hydrological optimized Informer model according to claim 5, characterized in that, In step S42, the Informer model is an improved Informer model, and the improvement method is specifically as follows: S421. Construct time feature encoding. Decompose the hydrological time feature in the hydrological spatio-temporal feature through EMD to obtain discrete IMF components and continuous IMF components. Use the sine function feature encoding method to generate the feature encoding at time d for the discrete IMF components: where Z is the time encoding method, is the sine function encoding under the Z time encoding method, per Z is the discrete IMF component, and d represents the preset period length under the Z time encoding method; Use the CountEncoder encoding method to generate the feature encoding at time d for the continuous IMF components; S422. Construct the longitude and latitude encoding of the spatial feature, and use the Gaussian kernel function mapping to encode the hydrological spatial feature in the hydrological spatio-temporal feature and encode the basin topological relationship: Among them, c lon , c lat respectively represent the longitude coordinate and the latitude coordinate of the center coordinates of the basin. lon i , lat i respectively represent the longitude coordinate and the latitude coordinate of the monitoring point. β represents the spatial attenuation coefficient, represents the spatial feature code of the x-th monitoring point, and the hydrological spatial features include the longitude and latitude coordinate vectors of the monitoring point and the key image frame features; S423. Construct the key image frame feature encoding of the spatial feature, and use the convolutional neural network to encode the key image frame features of the time series. This process can be expressed as: Among them, Y (H1+L1) is represented as a feature vector after H1 graph convolutions and L1 graph poolings, and are the associated features of the key image frame and its neighboring key image frames in the H1-th graph convolution respectively. Yconv(·) represents the graph convolution layer, ReLU represents the activation function, and Ypool(·) represents the graph pooling layer; S424. Use the multi-scale three-layer perceptron MLP to process the input time feature encoding and spatial feature encoding, and process them into a long-window time feature set L, a medium-window time feature set M, a short-window time feature set S, and a spatial feature set D respectively to obtain the final multi-scale perception fusion feature: D = MLP(E⊕H), Among them, E represents the hydrological spatial feature encoding in the hydrological spatio-temporal features, H represents the key image frame feature encoding in the hydrological spatial features, NOR represents the normalization algorithm, MLP represents the multi-layer perceptron operation, b1 represents the feature sets of different scales, ⊕ represents concatenation, and G′ LMSD represents the fused features after normalization, and G LMSD represents the final multi-scale perceptual fusion features. The spatial feature set D represents the final spatial embedding feature vector obtained by concatenating the hydrological spatial feature encoding E in the hydrological spatio-temporal features and the key image frame feature encoding H in the hydrological spatial features through the multi-scale perceptron MLP; S425. Input the obtained multi-scale perception fusion feature into the Informer model as the input feature encoding to obtain the improved Informer model. The improved Informer model also introduces a window sparse self-attention mechanism, a distillation mechanism, and a fully connected output mechanism. The output of the convolutional neural network is used to access the first input channel of the improved Informer model.

8. A surface water prediction method based on a hydrological optimized Informer model according to claim 7, characterized in that, The window sparse self-attention mechanism in step 425 includes: Use the sliding window form combined with the Fourier transform to characterize the self-attention mechanism; Use the sliding window form to limit the scope of action of the attention mechanism so that it only calculates the attention weight within the local window; Use the Fourier transform to convert the time series data from the time domain to the frequency domain, which can not only capture local time fluctuations but also process global trends.

9. A surface water prediction method based on a hydrological optimized Informer model according to claim 5, characterized in that, The calculation expression for performing model fusion according to the mean square error method is: f F = W I * f d + W T * f D , Among them, f d is the predicted value of the upstream surface water obtained by the Informer model, f D is the predicted value of the downstream surface water obtained by the GRU model, f F is the final predicted value after fusion, W I and W T are the weight coefficients of the Informer model and the GRU model respectively, θ I and θ L represent the mean square errors of the prediction results of the Informer model and the GRU model respectively.

10. A surface water prediction system based on a hydrological optimized Informer model, characterized in that, The system is used to execute the method described in any one of claims 1-9. The system includes a data collection module, a data preprocessing module, a feature extraction module, a model training module, and a model prediction module; The data collection module is used to collect multi-source data related to upstream and downstream monitoring points of surface water, and the data includes hydrological data, meteorological data, topographic data and water quality data; The data preprocessing module is used to preprocess the multi-source data to obtain preprocessed data; The preprocessing includes data cleaning, normalization processing and missing value filling; The feature extraction module is used to extract features from the preprocessed data to obtain water quality features and hydro-spatial-temporal features of upstream and downstream monitoring points, and construct cumulative dissolved oxygen saturation features. The hydro-spatial-temporal features include hydro-temporal features and hydro-spatial features; The model training module is used to input the water quality features, hydro-spatial-temporal features and dissolved oxygen saturation features of upstream and downstream into the multi-modal prediction model of the pre-trained improved Informer model to obtain the prediction results of upstream surface water and downstream surface water, and use the mean square error method to fuse the prediction results of upstream surface water and downstream surface water; The model prediction module is used to output the prediction results of surface water obtained by fusing the multi-modal prediction model of the improved Informer model, and the prediction results include changes in water volume, water quality and water level.

Images