A method, device and medium for predicting concentration of chlorophyll in a lake or reservoir

Abstract

The present application relates to a kind of lake reservoir type chlorophyll concentration prediction method, device and medium, method includes the following steps: obtaining the water quality, water dynamics and weather of the online automatic monitoring data of lake reservoir including forecast point and upstream point;The online automatic monitoring data is preprocessed, and the data after processing is obtained;The data after processing is carried out feature extraction to obtain the water quality feature and water dynamics index feature of forecast point and upstream point, and construct cumulative illumination feature;Upstream water quality feature, water dynamics index feature and illumination feature are input into the fusion prediction model pre-trained, to obtain upstream transport chlorophyll prediction result and time series influence chlorophyll prediction result, using error reciprocal method upstream transport chlorophyll prediction result and time series influence chlorophyll prediction result are fused;The chlorophyll prediction result obtained by fusion is output.Compared with prior art, the present application has the advantages of high accuracy, strong stability and the like.

Description

Technical Field

[0001] This invention relates to the field of water quality data prediction technology, and in particular to a method, device and medium for predicting chlorophyll concentration in lakes and reservoirs. Background Technology

[0002] Eutrophication refers to the imbalance of aquatic ecosystems in lakes and other bodies of water caused by excessive nutrient levels due to human activities. Eutrophication has become a major problem facing most lakes and reservoirs worldwide, representing a typical form of water pollution. Chlorophyll a is a water quality indicator that can characterize eutrophication and algal growth. Therefore, understanding the long-term spatiotemporal variations of chlorophyll a in lakes and reservoirs, revealing driving factors, and making predictions are of great significance for early warning of lake and reservoir risks and effective management of key areas.

[0003] Parametric statistics and ecodynamic models have long been traditional methods for modeling chlorophyll, and much research has been conducted on model prediction of chlorophyll a. However, existing models still have limited consideration of hydrodynamic conditions, often neglecting the continuous influence of environmental factors such as hydrodynamics and light on water movement and the distribution of substances in the water, resulting in prediction models that cannot accurately describe the changing trends of chlorophyll a. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of low accuracy in the prior art by providing a method, device and medium for predicting chlorophyll concentration in lakes and reservoirs.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for predicting chlorophyll concentration in lakes and reservoirs includes the following steps:

[0007] S1: Obtain online automatic monitoring data of water quality, hydrodynamics, and meteorology for lakes and reservoirs, including predicted locations and upstream locations;

[0008] S2: Preprocess the online automatic monitoring data to obtain the processed data;

[0009] S3: Extract features from the processed data to obtain water quality and hydrodynamic characteristics of the predicted and upstream locations, and construct cumulative illumination characteristics.

[0010] S4: Input the upstream water quality characteristics, hydrodynamic index characteristics and light characteristics into the pre-trained fusion prediction model to obtain the upstream transport chlorophyll prediction results and the time-series influence chlorophyll prediction results. Use the inverse error method to fuse the upstream transport chlorophyll prediction results and the time-series influence chlorophyll prediction results.

[0011] S5: The fusion prediction model outputs the chlorophyll prediction results obtained through fusion.

[0012] Furthermore, step S3 includes the following steps:

[0013] Randomly sample all water quality and hydrodynamic characteristics from the processed data to create a decision tree; calculate the regression prediction accuracy of data outside the sample bag. Features X of out-of-bag data i Apply random perturbations and calculate the regression prediction accuracy after perturbing the i-th feature. Calculate X i Importance P i The calculation expression is:

[0014]

[0015] In the formula, P i Let K represent the importance of the i-th feature, and K be the number of training samples.

[0016] For P i The feature importance ranking is obtained by sorting in descending order Q. i Then construct a fuzzy system; and sort the features Q. i Input the fuzzy system and obtain the final fuzzy system score for each feature. Select the features as input for the subsequent model based on the fuzzy system scores.

[0017] Furthermore, the method for calculating the cumulative illumination characteristics is as follows:

[0018] The illumination dataset is obtained by recording the cumulative illumination data in reverse order. The effect of turbidity on illumination at time r is calculated using the Lambert-Beer formula. The calculation expression is as follows:

[0019]

[0020] In the formula, TA r The effect of turbidity on intensity at time r, where ABS is the absorbance coefficient, and T... r Let r be the turbidity at time r;

[0021] The time weight at time r is calculated using the following expression:

[0022] TSr=eα*(r- l ag)

[0023] In the formula, TS r Here, lag represents the time weight, α represents the lag time (h), and α represents the time weight factor.

[0024] The cumulative illumination excluding lag time is calculated using the following expression:

[0025]

[0026] In the formula, AR represents cumulative illumination, and h represents the cumulative time period.

[0027] Furthermore, the fusion prediction model includes an Informer model and an LSTM model. The Informer model obtains the upstream transport chlorophyll prediction results based on the water quality characteristics and hydrodynamic index characteristics of the upstream points, while the LSTM model obtains the prediction results of the temporal influence of chlorophyll at the prediction points based on the water quality characteristics, hydrodynamic index characteristics, and cumulative light characteristics of the prediction points.

[0028] Furthermore, the specific steps in the training process of the pre-trained fusion prediction model include:

[0029] S401: Input the water quality and hydrodynamic parameters of the upstream points into Informe r The model yielded predictions for upstream chlorophyll transport.

[0030] S402: Input the water quality characteristics, hydrodynamic index characteristics and cumulative light characteristics of the prediction points into the LSTM model to obtain the time-series influence chlorophyll prediction results;

[0031] S403: Informe r The model and the LSTM model are used as base training models, and the models are fused according to the inverse error method to obtain the fused prediction model.

[0032] Furthermore, Informe r The model is the optimized Informe r The model and optimization method are as follows:

[0033] Step 4011: Construct time feature encoding. Decompose the time data according to different periods and generate the feature encoding for time d using the triangular feature encoding method:

[0034]

[0035] In the formula, Z represents the time decomposition method. Cosine feature encoding under Z-decomposition method, For sinusoidal feature encoding under Z-decomposition, per Z The period length under the Z-decomposition method;

[0036] Step 4012: Process the input temporal feature encoding using a multi-scale multilayer perceptron, dividing it into a long-time feature set L, a medium-time feature set M, and a short-time feature set S, respectively, to obtain the final multi-scale fused features:

[0037]

[0038] In the formula, NOR is the normalization algorithm, MLP is the multilayer perceptron operation, and c is the feature set at different scales. Indicates splicing, F′ LMS For the normalized fusion features, F LMS This represents the final fusion time characteristic;

[0039] Step 4013: Input the obtained fused time features into the Informer model as input feature encoding to obtain the optimized Informer model.

[0040] Furthermore, the LSTM model is an optimized LSTM model, and the optimization method is as follows:

[0041] Step 4021: Input feature T i Transformed into logarithmic return LT i ;

[0042] Step 4022: Optimize information acquisition in the LSTM algorithm using an attention mechanism;

[0043] Obtain LSTM processing LT i The hidden state matrix H obtained at that time is (h i-w , ..., h i-1 The temporal pattern matrix is ​​extracted using one-dimensional convolution, and the calculation expression is as follows:

[0044]

[0045] In the formula, C is the filter. Let w be the feature value extracted by the q-th filter at time p, w be the filter length, l be the padding length for convolution calculation, and W be the maximum weight length.

[0046] The attention formula is obtained by calculating the weights:

[0047]

[0048] In the formula, AT i Let σ be the attention vector, σ be the Sigmoid function, and WM be the weight matrix;

[0049] AT i with h i The final predicted value is obtained by summing the results after linear mapping;

[0050] Step 4023: Optimize the time step, learning rate, and filter length in the attention mechanism of the LSTM algorithm using the particle swarm optimization algorithm;

[0051] A particle swarm optimization algorithm is constructed by using timestep, learning rate lr, and filter length w to construct a dimensional space D, and randomly generating NP particles whose positions and velocities are represented by s. z and v z (i = 1, 2, ..., NP) represents the Nash efficiency coefficient NSE of the LSTM model prediction results, where the fitness function is:

[0052]

[0053] Where, x obs，z x is the actual value. model，z For predicted values, The mean of the actual values ​​is np, and the data size is np.

[0054] Based on the global optimal Y of all particles z and the local optimum Y of individual particle a za Update the velocity and position of each particle:

[0055] v za (t+1)=ωv za (t)+c1r1(Y za (t)-s za (t))+c2r2(Y z (t)-s za (t))

[0056] s za (t+1)=s za (t)+v za (t+1)

[0057] Where t is the number of iterations, ω is the inertia weight, c1 and c2 are learning factors, and r1 and r2 are random perturbation factors;

[0058] The particle swarm optimization algorithm is used to iterate until the termination condition is met, and the time step, learning rate and filter length at this point are output.

[0059] Step 4023: Input the logarithmic return rate, attention vector, time step, learning rate, and filter length into the LSTM model to obtain the prediction results.

[0060] Furthermore, the calculation expression for model fusion based on the inverse error method is as follows:

[0061] f F =w I *f I +w L *f L

[0062]

[0063] In the formula, f I f represents the predicted value of upstream transported chlorophyll obtained from the Informer model. L f represents the time-series influence of the chlorophyll prediction obtained from the LSTM model. F w represents the final predicted value after fusion. I w L The weight coefficients ε of the Informer model and the LSTM model are respectively. I ε L These represent the errors in the prediction results of the Informer model and the LSTM model, respectively.

[0064] In a second aspect, a lake-reservoir type chlorophyll concentration prediction device includes a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements any of the above-mentioned lake-reservoir type chlorophyll concentration prediction methods.

[0065] A third aspect of the invention is a storage medium having a program stored thereon, which, when executed, implements any of the above-mentioned methods for predicting chlorophyll concentration in lakes and reservoirs.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1) This invention preprocesses online automatic monitoring data of water quality, hydrodynamics, and light, and performs feature extraction and data cleaning to improve the accuracy of data prediction. By fusing prediction models, it obtains prediction results of upstream transported chlorophyll and prediction results of chlorophyll influencing time series, and realizes the analysis of the nonlinear relationship and time series influence of water quality, hydrodynamics, and light on chlorophyll distribution at upstream and prediction points, improves data stability, achieves rapid prediction, and can accurately predict chlorophyll data over a longer period of time.

[0068] 2) This invention integrates the Informer and LSTM models using the inverse error method. Based on the error of the model prediction results, different weights are combined to construct the INFORMER-LSTM prediction model, which effectively integrates multiple data points from upstream and prediction points, and combines the advantages of the two models, further improving the calculation speed and prediction accuracy of the prediction model.

[0069] 3) This invention proposes an accumulated illumination algorithm, which can more effectively construct features from illumination data. Attached Figure Description

[0070] Figure 1 This is a flowchart of the method of the present invention.

[0071] Figure 2 This is a flowchart of the INFORMER-LSTM fusion prediction model of the present invention.

[0072] Figure 3 This is a table comparing the prediction accuracy of the INFORMER-LSTM fusion prediction model of this invention with that of a single model and different fusion methods.

[0073] Figure 4 The image shows a comparison between the predicted chlorophyll a concentration after 1 hour and the actual chlorophyll a concentration after 1 hour, as shown in the example.

[0074] Figure 5 This is a list of prediction accuracies over a long period of time for the INFORMER-LSTM fusion prediction model in the example.

[0075] Figure 6 The image shows a comparison between the predicted chlorophyll a concentration 24 hours later and the actual chlorophyll a concentration 24 hours later, as shown in the example. Detailed Implementation

[0076] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0077] Example 1

[0078] This invention relates to a method, apparatus, and medium for predicting chlorophyll concentration in lakes and reservoirs, such as... Figure 1 As shown, the method includes the following steps:

[0079] S1: Obtain online automatic monitoring data of water quality, hydrodynamics, and meteorology for lakes and reservoirs, including predicted locations and upstream locations;

[0080] S2: Preprocess the online automatic monitoring data to obtain the processed data;

[0081] S3: Extract features from the processed data to obtain water quality and hydrodynamic characteristics of the predicted and upstream locations, and construct cumulative illumination characteristics.

[0082] S4: Input the upstream water quality characteristics, hydrodynamic index characteristics and light characteristics into the pre-trained fusion prediction model to obtain the upstream transport chlorophyll prediction results and the time-series influence chlorophyll prediction results. Use the inverse error method to fuse the upstream transport chlorophyll prediction results and the time-series influence chlorophyll prediction results.

[0083] S5: The fusion prediction model outputs the chlorophyll prediction results obtained through fusion.

[0084] Taking chlorophyll a as an example, the specific method is as follows:

[0085] S1: Obtain online automatic monitoring data of water quality, hydrodynamics, and meteorology for lakes and reservoirs, including predicted locations and upstream locations;

[0086] S2: Preprocess the online automatic monitoring data to obtain the processed data;

[0087] S201: Sort the data in ascending order based on time and remove duplicates;

[0088] S202: Preliminary filling was performed using linear regression;

[0089] S203: Perform data cleaning, classify and organize the acquired data according to data type; mine abnormal data features; use STL time series decomposition and density clustering methods to clean the data, and reconstruct abnormal data using the sum of trend components and periodic components after time series decomposition;

[0090] S204: Standardize the data, transforming it into a distribution with a mean of 0 and a standard deviation of 1.

[0091]

[0092] In the formula, X i For the data before processing, X i ′ represents the processed data, μ represents the average value of the index, and σ represents the standard deviation of the index.

[0093] Step 3: Select water quality and hydrodynamic indicators and construct illumination characteristics for the processed data;

[0094] The specific process for selecting water quality and hydrodynamic index characteristics from the processed data is as follows:

[0095] The original dataset was randomly sampled using the Bootstrap framework to obtain K training samples, each of which has N features.

[0096] Creating a decision tree Y k ;

[0097] Calculate the regression prediction accuracy for out-of-bag data on the training sample.

[0098] Then, the features X of the out-of-bag data were analyzed. i (i = 1, 2, ..., N) Apply random perturbation and calculate the regression prediction accuracy after perturbation.

[0099] Calculate X i Importance P of (i = 1, 2, ..., N) i :

[0100]

[0101] In the formula, P i Let K represent the importance of the i-th feature, and K be the number of training samples.

[0102] For P i The feature importance ranking is obtained by sorting in descending order Q. i Then, a fuzzy system is constructed, and a rule base is established; the sorted features Q are... i Input the fuzzy system and obtain the final fuzzy system score for each feature. Select the features as input for the subsequent model based on the fuzzy system scores.

[0103] Illumination features are constructed using a cumulative illumination algorithm, which is specifically as follows:

[0104] The illumination dataset is obtained by recording the cumulative illumination data in reverse order. The effect of turbidity on illumination at time r is then calculated using the Lambert-Beer formula.

[0105]

[0106] In the formula, TA r The effect of turbidity on intensity at time r, where ABS is the absorbance coefficient, and T... r Let r be the turbidity at time r;

[0107] Calculate the time weight at time r:

[0108] TSr=eα*(r- l ag)

[0109] In the formula, TS r Here, lag represents the time weight, α represents the lag time (h), and α represents the time weight factor.

[0110] Calculate the cumulative illumination excluding lag time:

[0111]

[0112] In the formula, AR represents cumulative illumination, and h represents the cumulative time period.

[0113] Step 4: Input water quality, hydrodynamic indicators, and illumination characteristics into Informe. r The INFORMER-LSTM fusion prediction model is trained with the LSTM prediction model to obtain the trained INFORMER-LSTM fusion prediction model.

[0114] Depend on Figure 2 As shown, water quality, hydrodynamic indicators, and illumination characteristics are input into Informe. r The specific steps for training an LSTM prediction model are as follows:

[0115] Step 401: Input upstream water quality characteristics and hydrodynamic characteristics into Informe r The model yielded predictions for upstream chlorophyll a transport.

[0116] Preferably, in step 401, Informe r The model is the optimized Informe r The model and optimization method are as follows:

[0117] Construct a time feature code by decomposing the time data into year, season, month, week, day, and hour, and using the triangular feature coding method to generate the feature code for time d:

[0118]

[0119] In the formula, Z represents the time decomposition method. Cosine feature encoding under Z-decomposition method, For sinusoidal feature encoding under Z-decomposition, per Z The period length is given by the Z-decomposition method.

[0120] The input time feature encoding is processed using a multi-scale multilayer perceptron, which consists of a long-term feature set L for processing yearly and seasonal features, a medium-term feature set M for processing monthly and weekly features, and a short-term feature set S for processing daily and hourly features, resulting in the final multi-scale fused features:

[0121]

[0122] In the formula, NOR is the normalization algorithm, MLP is the multilayer perceptron operation, and c is the feature set at different scales. Indicates splicing, F′ LMS For the normalized fusion features, F LMS This represents the final fusion time characteristic.

[0123] Input the obtained fusion time features into Informe r The model is used as input feature encoding to obtain the optimized Informer model.

[0124] Step 402: Input the water quality characteristics, hydrodynamic index characteristics and cumulative light characteristics of the predicted points into the LSTM model to obtain the prediction results of the time-series influence on chlorophyll a;

[0125] Preferably, the LSTM model is an optimized LSTM model, and the optimization method is as follows:

[0126] Transform input features into T i For the logarithmic return LT i :

[0127] LTi =ln(T) i+1 )-ln(T i )

[0128] Optimize information acquisition in the LSTM algorithm using an attention mechanism;

[0129] Obtain LSTM processing LT i The hidden state matrix H obtained at that time is (h i-w , ..., h i-1 The temporal pattern matrix is ​​extracted using one-dimensional convolution.

[0130]

[0131] In the formula, C is the filter. Let w be the feature value extracted by the q-th filter at time p, w be the filter length, l be the padding length for convolution calculation, and W be the maximum weight length.

[0132] The attention formula is obtained by calculating the weights:

[0133]

[0134] In the formula, AT i Let σ be the attention vector, σ be the sigmoid function, and WM be the weight matrix.

[0135] AT i with h i The final predicted value is obtained by summing the results after linear mapping.

[0136] The particle swarm optimization algorithm is used to optimize the time step, learning rate, and filter length in the attention mechanism of the LSTM algorithm.

[0137] A particle swarm optimization algorithm is constructed by using timestep, learning rate lr, and filter length w to construct a dimensional space D, and randomly generating NP particles whose positions and velocities are represented by s. z and v z (i = 1, 2, ..., NP) represents the fitness function, which is the Nash efficiency coefficient NSE of the LSTM model's prediction results.

[0138]

[0139] Where, x obs，z x is the actual value. model，z For predicted values, is the mean of the actual values, and np is the amount of data.

[0140] Based on the global optimal Y of all particles z and the local optimum Y of individual particle a zaUpdate the velocity and position of each particle:

[0141] v za (t+1)=ωv za (t)+c1r1(Y za (t)-s za (t))+c2r2(Y z (t)-s za (t))

[0142] s za (t+1)=s za (t)+v za (t+1)

[0143] Where t is the number of iterations, ω is the inertia weight, c1 and c2 are learning factors, and r1 and r2 are random perturbation factors.

[0144] The particle swarm optimization algorithm is used for iteration until the termination condition is met, and the time step, learning rate, and filter length at this point are output.

[0145] The logarithmic return rate, along with the attention vector, time step, learning rate, and filter length, are input into the LSTM model to obtain the prediction results.

[0146] Step 403: Informe r The model and the LSTM model are used as base training models, and the models are fused according to the inverse error method to obtain the INFORMER-LSTM fusion prediction model.

[0147] The Informer model and the LSTM model are used as base training models. The models are fused according to the inverse error method to obtain the INFORMER-LSTM fusion prediction model. The specific steps are as follows:

[0148] f F =w I *f I +w L *f L

[0149]

[0150]

[0151] In the formula, f I f represents the predicted value of upstream transported chlorophyll a obtained from the Informer model. L f represents the time-series influence of the predicted chlorophyll a value obtained from the LSTM model. F w represents the final predicted value after fusion. I w LThe weight coefficients ε of the Informer model and the LSTM model are respectively. I ε L These represent the errors in the prediction results of the Informer model and the LSTM model, respectively.

[0152] Step 5: Apply the INFORMER-LSTM fusion prediction model to predict the chlorophyll a concentration in the lake / reservoir type and obtain the prediction results.

[0153] A comparison of prediction accuracy results after fusing a single model with the inverse error method, such as... Figure 3 As shown, the INFORMER-LSTM fusion model demonstrates a significant improvement in prediction accuracy within a short period of one hour compared to single models and other fusion methods. Figure 4 As shown, this fusion model has a good effect in predicting chlorophyll a mutations.

[0154] like Figure 5 and Figure 6 As shown, this fusion model, by using the Informer model for fusion, solves the problem of data non-stationarity and also has a good fitting effect and high prediction accuracy over a long period of 24 hours.

[0155] This embodiment also includes a lake / reservoir type chlorophyll concentration prediction device, comprising a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the lake / reservoir type chlorophyll concentration prediction method provided by this invention. The program code for implementing the method of this invention can be written in any combination of one or more programming languages. This program code can be provided to the processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when the program code is executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server.

[0156] This embodiment also includes a storage medium storing a program thereon, which, when executed, implements the lake / reservoir type chlorophyll concentration prediction method provided by the present invention. The readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0157] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting chlorophyll concentration in lakes and reservoirs, characterized in that, Includes the following steps: S1: Obtain online automatic monitoring data of water quality, hydrodynamics, and meteorology for lakes and reservoirs, including predicted locations and upstream locations; S2: Preprocess the online automatic monitoring data to obtain the processed data; S3: Extract features from the processed data to obtain water quality and hydrodynamic characteristics of the predicted and upstream locations, and construct cumulative illumination characteristics. S4: Input the upstream water quality characteristics, hydrodynamic index characteristics and light characteristics into the pre-trained fusion prediction model to obtain the upstream transport chlorophyll prediction results and the time-series influence chlorophyll prediction results. Use the inverse error method to fuse the upstream transport chlorophyll prediction results and the time-series influence chlorophyll prediction results. S5: The fusion prediction model outputs the chlorophyll prediction results obtained through fusion; The fusion prediction model includes an Informer model and an LSTM model. The Informer model obtains the upstream transport chlorophyll prediction results based on the water quality characteristics and hydrodynamic index characteristics of the upstream points, while the LSTM model obtains the prediction results of the temporal influence of chlorophyll at the prediction points based on the water quality characteristics, hydrodynamic index characteristics, and cumulative light characteristics of the prediction points. The Informer model mentioned is an optimized Informer model, and the optimization method is as follows: Step 4011: Construct time feature codes by decomposing the time data according to different periods and generating them using triangular feature coding. Feature encoding at any given time; Step 4012: Use a multi-scale multilayer perceptron to process the input temporal feature encoding, which is processed into a long-time feature set L, a medium-time feature set M, and a short-time feature set S, respectively, to obtain the final multi-scale fused features; Step 4013: Input the obtained fused temporal features into the Informer model as input feature encoding to obtain the optimized Informer model; The LSTM model mentioned is an optimized LSTM model, and the optimization method is as follows: Step 4021: Input features Converted to logarithmic return ; Step 4022: Optimize information acquisition in the LSTM algorithm using an attention mechanism; Step 4023: Optimize the time step, learning rate, and filter length in the attention mechanism of the LSTM algorithm using the particle swarm optimization algorithm; Construct a particle swarm optimization algorithm, using timestep, learning rate (lr), and filter length. Construct a dimensional space D, and randomly generate NP particles whose positions and velocities are determined by... and (i=1,2,…,NP) represents the Nash efficiency coefficient NSE of the LSTM model prediction results, where the fitness function is: in, This is the actual value. For predicted values, The average of the actual values. For data volume; Based on the global optimality of all particles and individual particles Local optima Update the velocity and position of each particle: In the formula, The number of iterations. It is inertial weight. and As a learning factor, and For random disturbance factors; The particle swarm optimization algorithm is used for iteration until the termination condition is met, and the time step, learning rate, and filter length at this point are output.

2. The method for predicting chlorophyll concentration in lakes and reservoirs according to claim 1, characterized in that, Step S3 includes the following steps: Randomly sample all water quality and hydrodynamic characteristics from the processed data to create a decision tree; calculate the regression prediction accuracy of data outside the sample bag. Features of data outside the bag Apply a random perturbation and calculate the first... Regression prediction accuracy after feature perturbation ;calculate Importance The calculation expression is: In the formula, For the first The importance of each feature The number of training samples; right The feature importance sequence is obtained by sorting in descending order. Then construct a fuzzy system; and assign the feature importance sequence. Input the fuzzy system and obtain the final fuzzy system score for each feature. Select the features as input for the subsequent model based on the fuzzy system scores.

3. The method for predicting chlorophyll concentration in lakes and reservoirs according to claim 1, characterized in that, The method for calculating the cumulative illumination characteristics is as follows: The illumination dataset is obtained by recording the cumulative illumination data in reverse order. Turbidity is then calculated using the Lambert-Beer formula. r The effect of time on illumination can be calculated using the following expression: In the formula, for r The turbidity at any given time affects the intensity. The absorption coefficient is . for r Turbidity at any given moment; calculate r The time weight at each moment is calculated using the following expression: In the formula, As time weight, For the time lag, Time weighting factor; The cumulative illumination excluding lag time is calculated using the following expression: In the formula, To accumulate light, This is the cumulative time period.

4. The method for predicting chlorophyll concentration in lakes and reservoirs according to claim 1, characterized in that, The specific steps of training the pre-trained fusion prediction model include: S401: Input the water quality characteristics and hydrodynamic indicators of upstream points into the Informer model to obtain the upstream transport chlorophyll prediction results; S402: Input the water quality characteristics, hydrodynamic index characteristics and cumulative light characteristics of the prediction points into the LSTM model to obtain the time-series influence chlorophyll prediction results; S403: Use the Informer model and LSTM model as base training models, and fuse the models according to the inverse error method to obtain the fused prediction model.

5. The method for predicting chlorophyll concentration in lakes and reservoirs according to claim 1, characterized in that, The result obtained in step 4011 The feature encoding at time step is: In the formula, This is a time-decomposition method. for Cosine feature encoding under decomposition method for Sine feature encoding under decomposition method for Period length under the decomposition method; The multi-scale fusion features obtained in step 4012 are as follows: In the formula, NOR is the normalization algorithm, MLP is the multilayer perceptron operation, and c is the feature set at different scales. Indicates splicing, The normalized fusion characteristics This represents the final fusion time characteristic.

6. The method for predicting chlorophyll concentration in lakes and reservoirs according to claim 1, characterized in that, The specific steps in step 4022 to optimize the information acquisition of the LSTM algorithm using the attention mechanism include: Obtain LSTM processing The hidden state matrix obtained at that time =( ,⋯, The temporal pattern matrix is ​​extracted using one-dimensional convolution, and the calculation expression is as follows: In the formula, For filters, For the first Each filter in Feature values ​​extracted at each time step For filter length, Calculate the padding length for convolution. Indicates the maximum length of the weight; The attention formula is obtained by calculating the weights: In the formula, For attention vectors, For the Sigmoid function, This is the weight matrix; Will and The final predicted value is obtained by summing the results after linear mapping.

7. The method for predicting chlorophyll concentration in lakes and reservoirs according to claim 4, characterized in that, The calculation expression for model fusion based on the inverse error method is as follows: In the formula, These are the predicted values ​​of upstream transported chlorophyll obtained from the Informer model. The time-series influence of chlorophyll prediction values ​​obtained from the LSTM model. This is the final predicted value after fusion. , These are the weight coefficients for the Informer model and the LSTM model, respectively. , These represent the errors in the prediction results of the Informer model and the LSTM model, respectively.

8. A lake / reservoir type chlorophyll concentration prediction device, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements a method for predicting chlorophyll concentration in a lake or reservoir as described in any one of claims 1-7.

9. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements a method for predicting chlorophyll concentration in lakes and reservoirs as described in any one of claims 1-7.

Images