Lake and reservoir algal bloom prediction and early warning system and method based on PCA-RBF neural network and time-space fusion

Abstract

The invention provides a lake and reservoir algal bloom prediction and early warning system and method based on a PCA-RBF neural network and time-space fusion. Multi-dimensional water quality parameters are monitored and obtained in real time through a water quality sensor array device carried on an unmanned ship; key water quality parameters are dynamically screened by calculating Pearson's correlation coefficients of the water quality parameters, and PCC gt is reserved; a time sequence data set is generated according to the index of 0.22; inputting the time sequence data set into a PCA-RBF neural network prediction model for dynamic prediction, outputting a chlorophyll a concentration prediction value, and generating a prediction sequence; calculating the water bloom outbreak probability of the prediction sequence of the chlorophyll a concentration by adopting a time sequence decomposition and ARIMA combined algorithm; and carrying out risk grade division on the algal bloom outbreak probability according to a grading early warning rule. According to the method, dimension reduction is carried out on multi-dimensional water quality parameters through the PCA-RBF neural network, the concentration of chlorophyll a is predicted, a concentration time sequence rule is analyzed by adopting an STL-ARIMA combined algorithm, and graded early warning signals are output by fusing spatio-temporal characteristics, so that accurate prevention and control of water bloom risks are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of environmental monitoring, in particular to a lake and reservoir water bloom prediction and early warning method based on PCA-RBF neural network and space-time fusion. BACKGROUND

[0002] The current water quality prediction technology has significant defects in dealing with the high-dimensional, nonlinear and non-stationary characteristics of lake and reservoir water bodies due to the reliance on a single algorithm model. The main challenges are the insufficient adaptability of spatial dimension reduction methods and the poor stability of time prediction models. Current water quality monitoring usually requires the collection of 10+ parameters (such as TP, TN, COD, etc.). Due to the strong correlation between parameters, the model is overfitted, and the redundant information in high-dimensional data interferes with the extraction of effective features. Traditional models lack the ability to capture nonlinear relationships and cannot represent the threshold effect of dissolved oxygen on chlorophyll a. For example, standard PCA only retains the direction with the maximum global variance, ignoring the local nonlinear sensitive area (such as the DO threshold point), resulting in a key feature loss rate of >15%. When heavy rain runoff causes a 50% surge in TP concentration, PCA dimension reduction loses the nonlinear coupling effect of DO and TP, and the prediction error is 22% larger than that of the full-dimensional model. While feature selection algorithms such as random forest can filter important indicators, they cannot eliminate the multicollinearity between parameters, limiting the model's generalization ability. In terms of time prediction, a single ARIMA is used for non-stationary sequences (such as a 300% surge in Chl-a within 6 hours during an algae outbreak), which requires high-frequency differencing (d≥2), resulting in a prediction value standard deviation of >2.5 μg / L and a mutation point detection lag of more than 12 hours. Deep learning time series models such as LSTM have a parameter quantity of 1.2×10 7 (assuming 3 hidden layers with 256 nodes), the single inference delay on a Jetson Nano edge device is >500 ms, which cannot meet the minute-level high-frequency early warning requirements. SUMMARY

[0003] To overcome the shortcomings of the prior art, the present application proposes a lake and reservoir water bloom prediction and early warning system and method based on PCA-RBF neural network and space-time fusion. The PCA-RBF neural network is used to reduce the dimensionality of multi-dimensional water quality parameters and predict the concentration of chlorophyll a. Then, the STL-ARIMA joint algorithm is used to analyze the concentration time series, fuse the space-time features, and output a graded warning signal, achieving precise prevention and control of water bloom risk.

[0004] The technical solution of the present application is as follows: a lake and reservoir water bloom prediction and early warning system based on PCA-RBF neural network and space-time fusion, comprising a data acquisition module, a data preprocessing module, a PCA dimension reduction module, an RBF neural network prediction module, and a dynamic correction module. The data acquisition module acquires multi-dimensional water quality parameters in real time through a water quality sensor array device. The data preprocessing module screens the acquired multidimensional water quality parameters using the Pearson correlation coefficient (PCC), and then selects the water quality parameters related to chlorophyll a concentration. The obtained water quality parameters are then standardized by Z-score. The PCA dimensionality reduction module calculates the covariance matrix of the Z-score standardized data, performs eigenvalue decomposition and principal component selection, and realizes principal component extraction; RBF neural network prediction module, the input layer receives the principal component after dimensionality reduction, the hidden layer uses Gaussian radial basis function RBF as the activation function, the number of nodes is determined by K-means clustering, the output layer is a single-node linear output, and the chlorophyll a concentration is predicted. ; Dynamic correction module: uses a sliding window mechanism to set a time window and store the measured chlorophyll a concentration y within the window real and predicted chlorophyll a concentration , when the predicted chlorophyll a concentration in the window A correction is triggered when the mean absolute percentage error (MAPE) exceeds 15%.

[0005] Furthermore, the data acquisition module monitors and collects 16 water quality parameters in real time, including water temperature WT, pH value, dissolved oxygen DO, electrical conductivity EC, turbidity Turb, total phosphorus TP, total nitrogen TN, ammonia nitrogen NH3-N, permanganate index CODMn, chemical oxygen demand COD and chlorophyll aChl-a.

[0006] Furthermore, the data preprocessing module screens the collected water quality parameters through the Pearson correlation coefficient PCC formula (1), sets its threshold, and retains those with |PCC| > 0.22. Then, four parameters significantly correlated with chlorophyll a concentration are screened and Z-score normalized using formula (2). (1) Formula 1 represents the linear correlation between chlorophyll a and other water quality parameters, where X i is the i-th monitoring value of chlorophyll a, Y i is the i-th monitoring value of a water quality parameter, is the average value of chlorophyll a monitoring value, is the average value of the monitoring value of a certain water quality parameter; The four water quality parameters significantly correlated with chlorophyll a concentration were Z-score standardized using formula (2); (2) is the input data after water quality parameter standardization, μ train ,σ trainis the mean and standard deviation calculated based on the training set.

[0007] Further, the PCA dimension reduction module performs eigenvalue decomposition on the covariance matrix of the Z-score standardized data, (3) Z-score standardization of water quality parameters, Z-score standardization matrix of water quality parameters is the eigenvalue corresponding to the index, arranged in descending order; is the eigenvector corresponding to the index; Further principal component selection is performed to realize principal component extraction, and formula (4) outputs the principal component Z t . (4) W is a matrix composed of the eigenvectors of the first two principal components, which reduces the four-dimensional original data to three-dimensional principal components Z t by matrix multiplication, retaining 92% of the explained variance and realizing data dimension reduction and feature extraction.

[0008] Further, the RBF neural network prediction module inputs Z t into the RBF neural network to predict the concentration of chlorophyll a, and the RBF neural network structure formulas (5) and (6) are, The clustering center cj is clustered by K-means on the training set Z, the elbow rule determines the center number m=15, and the kernel width σ j is taken as 1.2 times the average distance of the jth sample to the nearest center, and the orthogonal least squares method OLS is used to solve the minimization Optimization of weights; (5) (6) is the predicted value of chlorophyll a concentration, w j is the weight from the hidden layer to the output layer, b is the bias, ϕ j (Z t ) is the output of the jth hidden layer node, which calculates the distance between the input data Z t and the jth clustering center c j , and converts it to weight through Gaussian radial function.

[0009] Further, the dynamic correction module sets the trigger condition in the RBF neural network dynamic correction mechanism, adopts the sliding window mechanism, sets the time window to 7 days, and stores the measured chlorophyll a concentration y realwith the predicted value When the windowed predicted value is triggered, the increment in the RBF neural network structure is updated, and the update only fine-tunes the output layer weight w j and the bias b, the cluster center c j and the kernel width σ j are kept unchanged, and the formula (7) and formula (8) (7) (8) wherein, is the predicted value of chlorophyll a concentration, y t is the measured value of chlorophyll a concentration, and MAPE is used to evaluate the deviation degree of the predicted value from the measured value; w new is the new weight matrix of the RBF neural network output layer, corresponding to the connection weight of the hidden layer node to the output layer, b new is the new bias term of the output layer, used to adjust the baseline offset of the predicted value, w old , b old is the old weight and bias before updating; is the hidden layer output matrix, with a dimension of 3x15, 3 being the sample number and 15 being the hidden layer node number, Φ T is the standardization matrix of the hidden layer output matrix; η is the learning rate, taking 0.01, to control the update step; is the core matrix operation of the OLS algorithm, used to solve the least square solution.

[0010] The lake bloom prediction and early warning method based on PCA-RBF neural network and spatiotemporal fusion has the following steps: S1, real-time monitoring and acquisition of multi-dimensional water quality parameters through a water quality sensor array device; S2, calculating the Pearson correlation coefficient PCC of the multi-dimensional water quality parameters obtained in S1, dynamically selecting water quality indicators, retaining the water quality parameters with |PCC|>0.22, and then selecting four parameters related to chlorophyll a concentration to generate a time series data set; S3, inputting the time series data set generated in S2 into the PCA-RBF neural network prediction model for dynamic prediction, outputting the predicted value of chlorophyll a concentration, and generating a predicted sequence of chlorophyll a; S4, calculating the bloom outbreak probability using the time series decomposition STL+ARIMA joint algorithm on the predicted sequence of chlorophyll a concentration; S5, dividing the bloom outbreak probability into risk levels according to the grading early warning rules; S6, the water bloom risk level is determined by comparing the grading early warning rules, and the water bloom risk level is obtained, at this time the unmanned ship carries out corresponding operation response.

[0011] Further, the water quality sensor array device in S1 is carried in the unmanned ship, the sampling frequency is 5 times / hour, and 1L of water sample is collected at 20-30m underwater.

[0012] Further, in S3, the predicted value of chlorophyll a concentration is output, when input Z t is input into the RBF neural network to obtain . As part of the input, the rolling prediction is performed to obtain a short-term prediction sequence of chlorophyll a concentration 1 hour.

[0013] Further, in S4, the water bloom outbreak probability is predicted by the prediction sequence of chlorophyll a concentration; Seasonal time series STL decomposition is performed, formula (9), LOESS local weighted regression is adopted, window = 24 hours, long-term change fitting is performed to obtain trend item T t , formula (10); periodicity is modeled by Fourier series to obtain seasonal item S t , formula (11); random fluctuations after removing trend and periodicity to obtain residual item R t , formula (12); (9) Wherein, the chlorophyll a concentration time series is decomposed into trend item T t , seasonal item S t and residual item R t ; the trend item Tt reflects the long-term change trend, the seasonal item S t reflects the periodic fluctuation, and the residual item R t represents the random fluctuation; (10) The trend item is fitted by LOESS local weighted regression, and the data of adjacent 12 time points are weighted and averaged by cubic kernel function ; h is the bandwidth, which is 12; the smoothness of local fitting is controlled; (11) The seasonal item is modeled by Fourier series, considering three periods: daily period P1=24h, weekly period P2=168h, and annual period P3=8760h; and The amplitudes of the cosine and sine components of the kth cycle are determined by least square fitting; (12) The residual term represents the random fluctuation part after removing the trend and seasonal effect, which is used for subsequent ARIMA model prediction; The residual is predicted by ARIMA model formula (13), and the residual sequence {Rt} is 1st order difference formula (15), and ARMA model is fitted Formula (16), and finally the predicted value is obtained by inverse difference Formula (17); (13) The residual sequence {Rt} is modeled by ARMA, where B is the lag operator, B k R t = R t −k, p is the autoregressive order, d is the difference order, usually d=1, and the stationarity is confirmed by ADF test, q is the moving average order, and are model parameters, j takes 1 to p, is a white noise sequence; Model likelihood value (14) Where AIC criterion is used to optimize the hyperparameters p, d, q of ARIMA model, The search space is p,q∈[0,5]; (15) Where the residual sequence {Rt} is 1st order difference, which is stationary, and is convenient for ARMA model fitting; (16) is the residual sequence after fitting difference, is the autoregressive coefficient, is the moving average coefficient, and the future prediction value of the residual is obtained by accumulating the difference prediction value of the future Δt time points ; (17) The water bloom outbreak probability formula (18) is calculated, and its threshold μ h =20μg / L, σ h =2.5μg / L historical data standard deviation, and σ h is expanded to 3.5 in summer algal outbreak period, and then the probability mapping is used to obtain the standard normal cumulative distribution function Φ(z) formula (19), and the concentration Probability of chlorophyll-a concentration exceeding eutrophication threshold of 20ug / L, (18) Probability of chlorophyll-a concentration exceeding eutrophication threshold of 20ug / L, h =20ug / L, σ h is the standard deviation of historical data, enlarged to 3.5 in summer algal bloom period, to measure the degree of concentration fluctuation; (19) Φ(z) is the standard normal cumulative distribution function, which standardizes the predicted concentration to the standard normal variable z and calculates the probability of exceeding the threshold.

[0014] The beneficial effects of the present application are: through PCA-RBF space modeling, the multiple collinearity of water quality parameters is eliminated, the mean square error (MSE) of chlorophyll-a concentration prediction is reduced, the seasonal fluctuation of concentration is captured by combining STL-ARIMA time series decomposition, and the accuracy of water bloom prediction and early warning is 91.3%; the sliding window incremental learning (window=7 days) is introduced, when the prediction mean absolute percentage error (MAPE) is greater than 15%, the weight online update is completed, the adaptive increase and decrease of parameters is realized, and the cross-water migration learning only needs one week of data to achieve 85% prediction accuracy; the early warning response closed loop is optimized, the three-level risk is divided based on the normal distribution probability mapping, and the limitation of fixed threshold is broken through. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is the flow chart of the lake and reservoir water bloom prediction and early warning method based on PCA-RBF neural network and space-time fusion of the present application; Figure 2 is the PCC thermodynamic diagram of 16 kinds of water quality parameters in the present application; Figure 3 is the PCA-RBF model structure diagram in the present application; Figure 4 is the structure schematic diagram of the unmanned ship in the present application; Figure 5 is the Pearson correlation coefficient of part of the water quality parameters in the present application; Figure 6 is the early warning response flow chart of the unmanned ship in the present application. DETAILED DESCRIPTION

[0016] In order to make the technical problems solved by the present application, the technical solutions adopted and the technical effects achieved more clear, the technical solutions of the embodiments of the present application will be further described in detail below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] Embodiments: Figures 1 to 5 As shown, the present application is based on a lake and reservoir water bloom prediction and early warning system based on PCA-RBF neural network and space-time fusion, including a data acquisition module, a data preprocessing module, a PCA dimension reduction module, an RBF neural network prediction module and a dynamic correction module. The data acquisition module acquires multi-dimensional water quality parameters in real time through a water quality sensor array device. The data preprocessing module screens the obtained multi-dimensional water quality parameters through the Pearson correlation coefficient PCC, and then screens out water quality parameters related to chlorophyll a concentration, and performs Z-score standardization on the obtained water quality parameters. The PCA dimension reduction module performs covariance matrix calculation on the Z-score standardized data, performs eigenvalue decomposition and principal component selection, and realizes principal component extraction. The RBF neural network prediction module receives the reduced principal components in the input layer, uses Gaussian radial basis function RBF as the activation function in the hidden layer, determines the number of nodes by K-means clustering, and performs single-node linear output in the output layer to predict the concentration of chlorophyll a. ; The dynamic correction module uses a sliding window mechanism, sets a time window, and stores the measured chlorophyll a concentration y real and the predicted value of chlorophyll a concentration When the predicted value of chlorophyll a concentration The average absolute percentage error MAPE exceeds 15% to trigger correction.

[0018] The method for predicting and warning lake and reservoir water bloom based on PCA-RBF neural network and space-time fusion uses a single unmanned boat platform for water sampling and detection, carries an automatic pipe retraction device for 30m water body stereoscopic sampling, and uses an NVIDIA Jetson AGX Orin (32GB RAM) edge computing unit, which is preinstalled with a PCA-RBF prediction module. The steps are as follows: S1, the unmanned ship water quality sensor array device, real-time monitoring of water temperature (WT), pH, dissolved oxygen (DO), conductivity (EC), turbidity (Turb), total phosphorus (TP), total nitrogen (TN), ammonia nitrogen (NH3-N), permanganate index (CODMn), chemical oxygen demand (COD), chlorophyll a (Chl-a) 16 indicators, sampling frequency 5 times / hour; S2, the multi-dimensional water quality parameters obtained in S1 are calculated to obtain the Pearson correlation coefficient PCC, the water quality indicators are dynamically screened, the water quality parameters with |PCC|>0.22 are retained, and four parameters related to the concentration of chlorophyll a are screened out, the PCC of the permanganate index (CODMn) is 0.72, the PCC of the total phosphorus (TP) is 0.23, the PCC of the total nitrogen (TN) is 0.28, and the PCC of the dissolved oxygen (DO) is 0.40, formula (1) (1) Wherein, PCC(X,Y) represents the linear correlation degree of chlorophyll a (X) and other water quality indicators (Y), X i is the i th monitoring value of chlorophyll a, is the average value of the monitoring value of chlorophyll a; Y i is the i th monitoring value of other water quality indicators, is the average value of the monitoring value of other water quality indicators.

[0019] The screened permanganate index (CODMn), total phosphorus (TP), total nitrogen (TN) and dissolved oxygen (DO) are used as four-dimensional input data Z = [CODMn, TP, TN, DO] of the model, and Z-score standardization formula (2) is performed to eliminate the dimension difference and prevent the large parameter from dominating the principal component direction; (2) Wherein, is the standardized input data, μ train ,σ train is the mean and standard deviation calculated based on the training set (fixed); (3) Wherein, water quality parameter Z-score standardization, water quality parameter Z-score standardization matrix, λ i is the eigenvalue corresponding to the index (arranged in descending order), v i is the corresponding eigenvector; (4) Wherein, QUOTE W is a matrix composed of the eigenvectors of the first two principal components. The four-dimensional original data is reduced to the three-dimensional principal component Z by matrix multiplication. t , retaining 92% of the explained variance, achieving data dimensionality reduction and feature extraction; S3, Z t Input into PCA-RBF neural network to predict chlorophyll a concentration, output chlorophyll a concentration prediction value, and generate chlorophyll a prediction sequence; In the RBF neural network structure formula (5) and formula (6), the cluster center c j The training set Z is clustered by K-means, and the number of centers m=15 and the kernel width σ are determined by the elbow rule. j Take 1.2 times the average distance from the j-th class sample to the nearest center (adaptively adjusted to optimize model fitting ability) and use the orthogonal least squares method OLS to minimize Optimize weights; (5) (6) is the predicted value of chlorophyll a concentration, w j is the weight from the hidden layer to the output layer, b is the bias, ϕ j (Z t ) is the output of the jth hidden layer node, calculate the input data Z t and the jth cluster center c j The distance is converted into weight through Gaussian radial function; The trigger condition is set in the RBF neural network dynamic correction mechanism, and the sliding window mechanism is used. The time window is set to 7 days, and the measured chlorophyll a concentration y in the storage window is stored. real and predicted value , when the predicted value in the window Triggered when the mean absolute percentage error MAPE>15%, the increment in the RBF neural network structure is updated, and the update only fine-tunes the output layer weight w j and bias b, maintain cluster center c j and kernel width σ j unchanged, formula (7) and formula (8) (7) (8) in, is the predicted value of chlorophyll a concentration, y t is the measured value of chlorophyll a concentration, and MAPE is used to evaluate the degree of deviation between the predicted value and the measured value; w newis the new weight matrix of the output layer of the RBF neural network, corresponding to the connection weight of the hidden layer nodes to the output layer, b new is the new bias term of the output layer, used to adjust the baseline offset of the predicted value, w old , b old is the old weight and bias before updating; is the hidden layer output matrix, with dimensions 3x15, 3 being the number of samples and 15 being the number of hidden layer nodes, Φ T is the standardization matrix of the hidden layer output matrix; η is the learning rate, taking 0.01, to control the update step size; is the core matrix operation of the OLS algorithm, used to solve the least squares solution; Input Z t into the RBF neural network to get , then as part of the input, to roll the prediction , to get the predicted sequence of chlorophyll a concentration , 1 hour.

[0020] S4, the predicted sequence of chlorophyll a concentration is calculated using the time series decomposition STL+ARIMA joint algorithm to calculate the probability of bloom outbreak; Seasonal time series STL decomposition, formula (9), using LOESS local weighted regression, window = 24 hours, fitting the long-term change to get the trend item T t formula (10); using Fourier series modeling periodicity to get the seasonal item S t formula (11); removing the trend and periodicity after the random fluctuation to get the residual item R t formula (12); (9) where the chlorophyll a concentration time series is decomposed into trend item T t , seasonal item S t and residual item R t ; trend item T t reflects the long-term trend, seasonal item S t reflects periodic fluctuations, and residual item R t represents random fluctuations; (10) Using LOESS local weighted regression to fit the trend item, through a cubic kernel function to weight average the data of the adjacent 12 time points; h is the bandwidth, taking 12; to control the smoothness of local fitting; (11) Seasonal terms are modeled by Fourier series, considering three periods: daily period P1=24h, weekly period P2=168h, yearly period P3=8760h;a k and b k are the amplitudes of the cosine and sine components of the kth period, respectively, determined by least square fitting; (12) Rtis the residual term, representing the random fluctuation part after removing the trend and seasonal effects, used for subsequent ARIMA model prediction; the residual sequence {Rt} is predicted by ARIMA model (13), and is first-order differenced (15), and fitted by ARIMA model (16), and finally the predicted value is obtained by inverse differencing (17); (13) ARIMA modeling is performed on the residual sequence {Rt}, where B is the lag operator, (B k R t = R t −k), p is the autoregressive order, d is the difference order, usually d=1, and stationarity is confirmed by ADF test, q is the moving average order, and are model parameters, j takes 1 to p, and {εt} is a white noise sequence; The model likelihood (14) is used to optimize the hyperparameters p, d, q of the ARIMA model, and the search space is p,q∈[0,5]; (15) where the residual sequence {Rt} is first-order differenced to make it stationary, facilitating ARMA model fitting; (16) is the fitted residual sequence after differencing, is the autoregressive coefficient, is the moving average coefficient, and the future prediction value of the residual is obtained by accumulating the difference prediction values of the future Δt time points ; (17) The water bloom outbreak probability formula (18) is calculated, and its threshold μ h =20μg / L (national standard GB 3838-2002 eutrophication threshold), σ h =2.5μg / L (based on historical data standard deviation, summer algal bloom period σ hextended to 3.5), then the standard normal cumulative distribution function Φ(z) formula (19) is obtained by using probability mapping, and the predicted concentration of chlorophyll a the probability of exceeding 20 μg / L, (18) for calculating the probability of chlorophyll a concentration exceeding the eutrophication threshold μ h =20μg / L at future time Δt, σ h is the standard deviation of historical data, and is extended to 3.5 in the summer algal bloom period, which is used to measure the degree of concentration fluctuation; (19) Φ(z) is the standard normal cumulative distribution function, which standardizes the predicted concentration to the standard normal variable z and calculates the probability of exceeding the threshold.

[0021] S5, the risk level of the algal bloom outbreak probability is divided according to the grading early warning rule; S6, the algal bloom risk level is determined by comparing the grading early warning rule, and the algal bloom risk level is obtained. At this time, the unmanned ship carries out corresponding operation response.

[0022] As Figure 6 shown, based on daily inspection operation, the unmanned ship reaches the sampling point, collects and detects water quality according to the requirements, judges the algal bloom risk level, if the risk level is green, carries out normal monitoring, if the risk level is yellow, needs to leave sample for inspection verification, if the risk level is red, the unmanned ship adds oxygen and adds algae-removing agent.

Claims

1. The lake and reservoir algal bloom prediction and warning system based on PCA-RBF neural network and spatiotemporal fusion is characterized by: It includes data acquisition module, data preprocessing module, PCA dimension reduction module, RBF neural network prediction module and dynamic correction module; The data acquisition module acquires multi-dimensional water quality parameters through real-time monitoring of the water quality sensor array device; The data preprocessing module analyzes the acquired multidimensional water quality parameters through Pearson correlation coefficient (PCC) analysis, selects water quality parameters related to chlorophyll a concentration, and performs Z-score standardization on the obtained water quality parameters; The PCA dimensionality reduction module calculates the covariance matrix of the Z-score standardized data, performs eigenvalue decomposition and principal component selection, and realizes principal component extraction; RBF neural network prediction module, the input layer receives the principal component after dimensionality reduction, the hidden layer uses Gaussian radial basis function RBF as the activation function, the number of nodes is determined by K-means clustering, the output layer is a single-node linear output, and the chlorophyll a concentration is predicted. ; Dynamic correction module: uses a sliding window mechanism to set a time window and store the measured chlorophyll a concentration y within the window real and predicted chlorophyll a concentration , when the predicted chlorophyll a concentration in the window A correction is triggered when the mean absolute percentage error (MAPE) exceeds 15%.

2. The lake and reservoir algal bloom prediction and early warning system based on PCA-RBF neural network and spatiotemporal fusion according to claim 1 is characterized in that: The data acquisition module monitors and collects 16 water quality parameters in real time, including water temperature WT, pH value, dissolved oxygen DO, conductivity EC, turbidity Turb, total phosphorus TP, total nitrogen TN, ammonia nitrogen NH3-N, permanganate index CODMn, chemical oxygen demand COD and chlorophyll a.

3. The lake and reservoir algal bloom prediction and early warning system based on PCA-RBF neural network and spatiotemporal fusion according to claim 1 is characterized in that The data preprocessing module filters the collected water quality parameters through the Pearson correlation coefficient PCC formula (1), sets the threshold |PCC| > 0.22, and retains the water quality parameters when the Pearson correlation coefficient analysis is greater than 0.

22. The retained water quality parameters are significantly correlated with the chlorophyll a concentration and are subjected to Z-score standardization, formula (2); （1） Formula (1) represents the linear correlation between chlorophyll a and other water quality parameters, where X i is the i-th monitoring value of chlorophyll a, Y i is the i-th monitoring value of a water quality parameter, is the average value of chlorophyll a monitoring value, is the average value of the monitoring value of a certain water quality parameter; Water quality parameters significantly correlated with chlorophyll a concentration were Z-score standardized using formula (2); （2） is the input data after water quality parameter standardization, μ train ,σ train are the mean and standard deviation calculated based on the training set.

4. The lake and reservoir algal bloom prediction and early warning system based on PCA-RBF neural network and spatiotemporal fusion according to claim 1 is characterized in that The PCA dimension reduction module uses the covariance matrix of the water quality parameter Z-score standardized data that is significantly correlated with the chlorophyll a concentration to perform eigenvalue decomposition formula (3). （3） Z-score standardization of water quality parameters, Water quality parameter Z-score standardized matrix, λ i are the eigenvalues ​​of the corresponding indicators, arranged in descending order; is the corresponding eigenvector; Then perform principal component selection to realize principal component extraction, formula (4), and reduce the dimension to three-dimensional principal component Z t ; （4） W is a matrix composed of the eigenvectors of the first two principal components. The four-dimensional original data is reduced to the three-dimensional principal component Z by matrix multiplication. t , retaining 92% of the explained variance, achieving data dimensionality reduction and feature extraction.

5. The lake and reservoir algal bloom prediction and early warning system based on PCA-RBF neural network and spatiotemporal fusion according to claim 1 is characterized in that The RBF neural network prediction module, Z t Input into RBF neural network to predict chlorophyll a concentration, RBF neural network structure formula (5) and formula (6), cluster center c j The training set Z is clustered by K-means, and the number of centers m=15 and the kernel width σ are determined by the elbow rule. j Take 1.2 times the average distance from the jth class sample to the nearest center and use the orthogonal least squares method OLS to minimize Optimize weights; （5） （6） is the predicted value of chlorophyll a concentration, w j is the weight from the hidden layer to the output layer, b is the bias, ϕ j (Z t ) is the output of the jth hidden layer node, and the input data Z is calculated t and the jth cluster center c j The distance is converted into a weight through a Gaussian radial function.

6. The lake and reservoir algal bloom prediction and early warning system based on PCA-RBF neural network and spatiotemporal fusion according to claim 1 is characterized in that The dynamic correction module sets the trigger condition in the RBF neural network dynamic correction mechanism, adopts the sliding window mechanism, sets the time window to 7 days, and stores the measured chlorophyll a concentration y in the window. real and predicted chlorophyll a concentration , when the predicted value in the window Triggered when the mean absolute percentage error MAPE>15%, the increment in the RBF neural network structure is updated, and the update only fine-tunes the output layer weight w j and bias b, maintain cluster center c j and kernel width σ j unchanged, formula (7) and formula (8) （7） （8） in, is the predicted value of chlorophyll a concentration, y t is the measured value of chlorophyll a concentration, and MAPE is used to evaluate the degree of deviation between the predicted value and the measured value; w new is the new weight matrix of the RBF neural network output layer, corresponding to the connection weight from the hidden layer nodes to the output layer, b new is the new bias term of the output layer, which is used to adjust the baseline offset of the predicted value, w old 、b old are the old weights and biases before updating; is the hidden layer output matrix, with a dimension of 3×15, 3 is the number of samples, 15 is the number of hidden layer nodes, Φ T is the normalized matrix of the hidden layer output matrix; η is the learning rate, which is 0.01, and controls the update step size; It is the core matrix operation of the OLS algorithm, used to solve the least squares solution.

7. A lake and reservoir algal bloom prediction and warning method based on a PCA-RBF neural network and spatiotemporal fusion, comprising the lake and reservoir algal bloom prediction and warning system based on a PCA-RBF neural network and spatiotemporal fusion according to any one of claims 1 to 6, comprising the following steps: S1, real-time monitoring and acquisition of multi-dimensional water quality parameters through a water quality sensor array device; S2, calculates the Pearson correlation coefficient PCC of the multidimensional water quality parameters obtained in S1, dynamically screens water quality indicators, retains water quality parameters with |PCC| > 0.22, and then screens out four parameters related to chlorophyll a concentration to generate a time series data set; S3, inputs the time series data set generated in S2 into the RBF neural network prediction model for dynamic prediction, outputs the predicted value of chlorophyll a concentration, and generates a prediction sequence of chlorophyll a; S4, the predicted series of chlorophyll a concentration is used to calculate the probability of algal bloom outbreak using the time series decomposition STL+ARIMA joint algorithm; S5, classify the risk level of algal bloom outbreak probability according to the graded warning rules; S6, compare the graded warning rules to determine the algal bloom risk level, and obtain the algal bloom risk level. At this time, the unmanned boat performs corresponding operational responses.

8. The lake and reservoir algal bloom prediction and early warning method based on PCA-RBF neural network and spatiotemporal fusion according to claim 7 is characterized in that: The water quality sensor array device in S1 is carried in the unmanned boat, with a sampling frequency of 5 times / hour, and collects 1L of water sample 20m underwater.

9. The lake and reservoir algal bloom prediction and early warning method based on PCA-RBF neural network and spatiotemporal fusion according to claim 7, characterized in that: S3 outputs the predicted value of chlorophyll a concentration. When Z is input t Input into RBF neural network to get , then As part of the input, rolling forecast , and obtain the short-term prediction series of chlorophyll a concentration 1 hour.

10. The lake and reservoir algal bloom prediction and early warning method based on PCA-RBF neural network and spatiotemporal fusion according to claim 7, characterized in that: In S4, the probability of algal bloom outbreak is predicted by the predicted sequence of chlorophyll a concentration; Perform seasonal time series STL decomposition, formula (9), use LOESS local weighted regression, window = 24 hours, fit the long-term changes to obtain the trend term T t Formula (10); Using Fourier series to model the periodicity, we get the seasonal term S t Formula (11): After removing the trend and cycle, the random fluctuations are obtained as the residual term R t Formula (12); （9） Among them, the chlorophyll a concentration time series Decomposed into trend term T t 、Seasonal Item S t and the residual term R t ; The trend term Tt reflects the long-term trend, and the seasonal term S t Reflecting cyclical fluctuations, the residual term R t represents random fluctuations; （10） The LOESS local weighted regression is used to fit the trend term, and the cubic kernel function is used to Perform weighted averaging on the data of 12 adjacent time points; h is the bandwidth, which is set to 12; controls the smoothness of the local fitting; （11） Fourier series was used to model seasonal terms, and three cycles were considered: daily cycle P1 = 24 h, weekly cycle P2 = 168 h, and annual cycle P3 = 8760 h; and Determined by least squares fitting, they represent the amplitudes of the cosine and sine components of the kth period, respectively; （12） The residual term is the random fluctuation part after removing the trend and seasonal effects, which is used for subsequent ARIMA model prediction; Use the ARIMA model formula (13) to predict the residuals, perform the first-order difference formula (15) on the residual series {Rt}, and fit the ARMA model Formula (16), and finally the predicted value is obtained by the reverse difference Formula (17); （13） Perform ARIMA modeling on the residual series {Rt}, where B is the lag operator, B k R t = R t −k, p is the autoregressive order, d is the difference order, usually d=1, and the stationarity is confirmed by the ADF test, q is the moving average order, and is the model parameter, j ranges from 1 to p, is a white noise sequence; Model likelihood (14) The AIC criterion is used to optimize the hyperparameters p, d, and q of the ARIMA model. The search space is p,q∈[0,5]; （15） Among them, the residual sequence {Rt} is subjected to first-order difference to make it stable, which is convenient for fitting the ARMA model; （16） is the residual sequence after fitting the difference, is the autoregressive coefficient, is the moving average coefficient, and the future prediction value of the residual is obtained by accumulating the differential prediction value of the future Δt time points ; （17） Formula (18) is used to calculate the probability of algal bloom outbreak, and its threshold μ is set. h =20μg / L, σ h =2.5μg / L standard deviation of historical data, summer algae bloom period σ h Expand to 3.5, and then use probability mapping to obtain the standard normal cumulative distribution function Φ(z) formula (19) to predict the concentration The probability of exceeding 20μg / L, （18） Used to calculate the chlorophyll a concentration exceeding the eutrophication threshold μ at the next Δt moment h = the probability of 20 μg / L, σ h is the standard deviation of historical data, which is expanded to 3.5 during the summer algae bloom period and is used to measure the degree of concentration fluctuation; （19） Φ(z) is the standard normal cumulative distribution function, which normalizes the predicted concentration to the standard normal variable z and calculates the probability that it exceeds the threshold.

Images