A long-time-scale drought index prediction method and system based on Prophet-GRU-CNN

Abstract

The application discloses a long-time-scale drought index prediction method and system based on Prophet-GRU-CNN, which comprises the following steps: obtaining a prediction factor with high correlation with an SPEI drought index; processing time series in a set time to form a data set, selecting the SPEI drought index as a target label value, and inputting the target label value into a GRU model to extract time characteristics; inputting the characteristics into a CNN model to obtain deep local characteristics; capturing influence factors in time series data to extract time characteristics, integrating the time characteristics into a Prophet model for prediction, and extracting time characteristics in historical SPEI data; dividing the data set into a training set and a test set according to a preset ratio, and evaluating the performance of the long-time-scale drought index prediction model. The application solves the problems of insufficient data, deep feature extraction and combination of multiple algorithms, supplements the data volume, strengthens the absorption of data time and space characteristics, improves the utilization rate of historical data, and improves the prediction effect of the long-time-scale drought index.

Description

Technical Field

[0001] This invention relates to the field of drought prediction, specifically to a long-term drought index prediction method and system based on Prophet-GRU-CNN. Background Technology

[0002] Currently, meteorological drought assessment primarily relies on rainfall as the main indicator to define the occurrence and severity of meteorological drought. Multifactor indicators (such as the Palmer index, Z-index, and SPEI) combine rainfall with other meteorological conditions to provide a more comprehensive assessment of drought conditions. Among them, the SPEI, due to its comprehensive consideration of potential evapotranspiration and precipitation, excels in continuous record-keeping and is widely used in drought research.

[0003] With the advancement of machine learning and deep learning technologies, in recent years, LSTM (Long Short-Term Memory) network models based on CEEMD (Complementary Ensemble Empirical Mode Decomposition) and ARIMA (Autoregressive Integrated Moving Average) models have been used for drought prediction. A VMD-GRU drought prediction model has been established by combining VMD (Variational Mode Decomposition) and GRU (Gated Recurrent Unit). Furthermore, an improved CQPSO (Chaotic Quantum Particle Swarm Optimization) algorithm has been used to optimize the hyperparameters of the GRU neural network, resulting in the VMD-CQPSO-GRU intelligent prediction model, which can be used for drought prediction. However, the performance of these models may be affected by data characteristics and model parameter selection. The effectiveness of the optimization algorithm is highly dependent on parameter selection and the local optima encountered during the optimization process. As the prediction time span increases, the model may face problems such as data sparsity, prediction uncertainty, and decreased model generalization ability. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a long-term drought index prediction method and system based on Prophet-GRU-CNN. By constructing the Prophet-GRU-CNN drought index prediction model, the drought index can be obtained. This not only improves the utilization rate of historical data, but also mines the deep mapping between the predictor and the label value. The combination of the three will greatly improve the prediction effect of the drought index.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] This invention proposes a long-term drought index prediction method based on Prophet-GRU-CNN, comprising:

[0007] S1. Correlation analysis was performed on meteorological factor data and SPEI (Standardized Precipitation Evapotranspiration Index) drought index at various scales to obtain predictive factors with high correlation to SPEI drought index.

[0008] S2. For the predictor factors in step S1, select time series within a set time period, preprocess them to form a dataset. Select the SPEI drought index as the target label value, input the predictor factor time series into the GRU model, and extract time features.

[0009] S3. Input the extracted temporal features into the CNN (Convolutional Neural Network) model to obtain deep local features. Input these features into the fully connected neural network layer used to predict the target data and output the spatiotemporal relationship between the predictor sequence and the target data.

[0010] S4. Obtain historical SPEI drought index data and time scales for relevant time scales, capture influencing factors in time series data to extract time features, integrate these features into the Prophet model for prediction, and extract time features from historical SPEI data.

[0011] S5. Based on steps S2-S4, obtain the Prophet-GRU-CNN model, divide the dataset in step S2 into training set and test set according to a preset ratio, and evaluate the performance of the Prophet-GRU-CNN model.

[0012] S6. Input the sample to be tested into the Prophet-GRU-CNN model to obtain the SPEI drought index prediction results.

[0013] Furthermore, in step S1, obtaining the predictor includes the following:

[0014] Predictive factors were selected, including but not limited to precipitation, 10m wind speed, relative humidity, peak logarithmic solar radiation, evaporation, runoff, convective potential energy, and net solar radiation.

[0015] Pearson correlation analysis was used to analyze the SPEI drought index, and the correlation coefficient between each meteorological data point and the SPEI drought index at each station was measured. The specific formula is as follows:

[0016]

[0017] Where, x i and y i This represents the i-th pair of observations in the sample data. This represents the average value of x. Let y represent the average value of y, and r represent the correlation coefficient.

[0018] Furthermore, in step S2, extracting time features includes the following sub-steps:

[0019] S201. Clean the time series data, impute missing values, integrate the data, and normalize the data to obtain a dataset of the corresponding dimensions.

[0020] S202 and GRU employ a gating mechanism, merging the forget gate and input gate into a single update gate with forgetting functionality. GRU also integrates cell states and hidden states, effectively addressing issues such as long-term memory loss and gradient vanishing during backpropagation in Long Short-Term Memory (LSTM) networks. GRU consists of a reset gate and an update gate.

[0021] Using the GRU model, temporal features are extracted by controlling the update and reset gates. The specific formula is as follows:

[0022] r t =σ(w xr x t +w hr h t-1 +b r )

[0023] z t =σ(w xz x t +w hz h t-1 +b z )

[0024]

[0025]

[0026] Where, r t Indicates resetting the gate output, z t This indicates an update to the gate output. h represents the candidate state at the current moment. t h represents the final output. t-1 Let x represent the hidden state at time t-1.t Let represent the input at time t, σ represent the sigmoid activation function, and w xr and w xz They represent x respectively t With r t and z t The weight relationship between them, w hr and w hz They represent h respectively t-1 With r t and z t The weight relationship between them, b r Indicates r t The corresponding bias term, b z Indicate z t The corresponding bias term, w xh and w hh They represent With x t and h t-1 The weight relationship between them, b h express The corresponding bias term, This represents the element-wise product of vectors.

[0027] Furthermore, in step S3, the spatiotemporal relationship between the output predictor sequence and the target data includes the following:

[0028] The extracted temporal features are fed into the CNN module, where convolutional and pooling layers in the two-layer CNN network extract deep local features of the temporal features. The feature matrix is ​​then output through a fully connected layer to obtain the spatiotemporal relationship between the predictor sequence and the target data.

[0029] Furthermore, in step S4, the specific formula for extracting the time features from the SPEI historical data is as follows:

[0030] y(t) = g(t) + s(t) + h(t) + ε(t)

[0031] Where g(t) represents the trend of the time series in non-periodic terms; s(t) represents the periodic term, or seasonal term, which is generally in units of weeks or years; h(t) represents the holiday term, which represents the impact of potential non-fixed periodic holidays in the time series on the predicted value; ε(t) represents the error term or residual term, which represents the fluctuations not predicted by the model and follows a Gaussian distribution.

[0032] Influencing factors include, but are not limited to, trends, cyclical factors, and the impact of special holidays.

[0033] Furthermore, in step S5, the prediction performance of the Prophet-GRU-CNN model is tested using the mean absolute error, root mean square error, Nash efficiency coefficient, and mean absolute percentage error. The specific formula is as follows:

[0034]

[0035]

[0036]

[0037]

[0038] Where MAE represents mean absolute error, RMSE represents root mean square error, NSE represents Nash efficiency coefficient, T represents time, MAPE represents mean absolute percentage error, m represents the number of data points, and x represents the mean absolute percentage error. i Represents the observed value. Indicates the predicted value; Represents the observed value at time t. Represents the simulated value at time t. This represents the average value of the observed values.

[0039] The closer MAE and RMSE are to 0, the better the prediction effect; the closer NSE is to 1, the stronger the fit between the two, that is, the higher the reliability of the model; the larger the MAPE result, the greater the deviation between the predicted value and the true value, that is, the worse the prediction effect of the model.

[0040] Furthermore, this invention also proposes a long-term drought index prediction system based on Prophet-GRU-CNN, characterized by including...

[0041] The predictor acquisition module is used to perform correlation analysis between meteorological factor data and the standardized precipitation evapotranspiration drought index at various scales, and to obtain predictors that have a correlation with the standardized precipitation evapotranspiration drought index greater than a set threshold.

[0042] The time feature extraction module is used to select time series within a set time period based on the predictor factors in the predictor acquisition module, preprocess them to form a dataset; select the standardized precipitation evapotranspiration drought index as the target label value, input the predictor time series into the gated recurrent unit model, and extract time features.

[0043] The module for obtaining the spatiotemporal relationship between the predictor sequence and the target data is used to input the extracted temporal features into the convolutional neural network model, obtain deep local features, and output the spatiotemporal relationship between the predictor sequence and the target data.

[0044] The historical data time feature extraction module is used to obtain historical data of standardized precipitation evapotranspiration drought index and the time scale of the data at relevant time scales, capture the influencing factors in the time series data to extract time features, integrate the features into the Prophet model for prediction, and extract the time features in the standardized precipitation evapotranspiration historical data.

[0045] The model evaluation module is used to divide the dataset in the time feature extraction module into training and test sets according to a preset ratio to evaluate the performance of the long-term drought index prediction model.

[0046] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the long-term drought index prediction method based on Prophet-GRU-CNN described above.

[0047] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program, which is executed by a processor to perform the long-term drought index prediction method based on Prophet-GRU-CNN.

[0048] The present invention adopts the above technical solution, and its significant technical effects compared with the prior art are as follows:

[0049] (1) No additional hardware is required.

[0050] (2) Based on the correlation analysis results, predictors of the SPEI drought index with high correlation were selected, which helps to simplify the model and improve the prediction performance.

[0051] (3) A Prophet-GRU-CNN model was constructed, which combines Prophet, GRU and CNN algorithms for predicting the SPEI drought index. This model utilizes different types of data and time series features to improve the prediction effect.

[0052] (4) When the dataset is limited, or even small, the deep learning model Prophet-GRU-CNN can improve prediction performance by utilizing complex feature extraction and spatiotemporal relationship modeling. It can better capture hidden patterns in the data, helping to compensate for insufficient data. Furthermore, introducing new meteorological features as full features can still provide useful information and improve the model's prediction performance. This introduction of full features can compensate for the problem of insufficient data. Attached Figure Description

[0053] Figure 1This is a flowchart illustrating the overall implementation of the present invention.

[0054] Figure 2 This is a correlation analysis diagram between various meteorological factors and the drought index in an embodiment of the present invention.

[0055] Figure 3 This is a network structure diagram of the Prophet-GRU-CNN model in an embodiment of the present invention.

[0056] Figure 4 This is a comparison chart of four model evaluation indicators for different models in the embodiments of the present invention.

[0057] Figure 5 This is a comparison chart of the predicted sequences of the five models in this embodiment of the invention on different time scales of the SPEI index. Detailed Implementation

[0058] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0059] This invention proposes a long-term drought index prediction method based on Prophet-GRU-CNN, such as... Figure 1 As shown, it includes the following steps:

[0060] S1. Obtain meteorological data from the National Climate Center and conduct correlation analysis between meteorological factor data and the standardized precipitation evapotranspiration drought indexes (SPEI01, SPEI03, SPEI06, and SPEI12) to obtain predictive factors with a correlation greater than 0.5 with the standardized precipitation evapotranspiration drought index. Prepare meteorological data from 35 meteorological stations and drought index data for SPEI01, SPEI03, SPEI06, and SPEI12, including precipitation, 10m wind speed, relative humidity, peak sunshine duration, evaporation, runoff, convective potential energy, and net solar radiation.

[0061] The correlation coefficients between each meteorological data point and the SPEI drought index were calculated using Pearson correlation analysis. The correlation coefficients between each meteorological data point and the standardized precipitation evapotranspiration drought index at each station were measured. The specific formula is as follows:

[0062]

[0063] Where, x i and y i This represents the i-th pair of observations in the sample data. This represents the average value of x. Let y represent the average value of y, and r represent the correlation coefficient.

[0064] The correlation analysis diagram between various meteorological factors and the drought index is shown in the figure below. Figure 2 As shown, where Figure 2 a, Figure 2 b, Figure 2 c and Figure 2 d represents the analysis results for the SPEI01, SPEI03, SPEI06, and SPEI12 drought indices, respectively. The images show that the absolute values ​​of the correlation coefficients between precipitation, 10m wind speed, relative humidity, peak sunshine hours, evaporation, runoff, convective potential energy, and net solar radiation and each SPEI drought index are relatively large. Therefore, the above climate data were selected as predictors of the SPEI drought index. Furthermore, it was found that as the time scale increases, the correlation coefficient series of each meteorological factor generally shifts towards the 0 axis. This is because the long-term SPEI drought index is more affected by climate anomalies such as El Niño and ocean current disturbances, and therefore its correlation with conventional meteorological data is lower than that of the short-term SPEI drought index.

[0065] The construction process of the Prophet-GRU-CNN model is as follows: Figure 3 As shown:

[0066] S2. Based on the predictive factors in step S1, select time series data from the past six months, preprocess them to form a dataset, including 7 series of meteorological data for the predictive factors selected in step S1 and 1 series of historical data for the SPEI drought index. The standardized precipitation-evapotranspiration drought index for the next month is selected as the target label value. The predictive factor time series are input into a gated recurrent unit model to extract temporal features, specifically:

[0067] S201. Clean the time series data, impute missing values, integrate the data, and normalize the data to obtain a dataset with a dimension of 8×6.

[0068] S202. Using the gated loop unit model, time features are extracted by controlling the update and reset gates. The specific formula is as follows:

[0069] r t =σ(w xr x t +w hr h t-1 +b r )

[0070] z t =σ(w xz x t +w hz ht-1 +b z )

[0071]

[0072]

[0073] Where, r t Indicates resetting the gate output, z t This indicates an update to the gate output. h represents the candidate state at the current moment. t h represents the final output. t-1 Let x represent the hidden state at time t-1. t Let represent the input at time t, σ represent the sigmoid activation function, and w xr and w xz They represent x respectively t With r t and z t The weight relationship between them, w hr and w hz They represent h respectively t-1 With r t and z t The weight relationship between them, b r Indicates r t The corresponding bias term, b z Indicate z t The corresponding bias term, w xh and w hh They represent With x t and h t-1 The weight relationship between them, b h express The corresponding bias term, This represents the element-wise product of vectors.

[0074] S3. After the GRU module, add a CNN layer to capture deep local features. The CNN layer can capture local features at different time steps and spatial locations through convolutional operations. The extracted temporal features are input into the convolutional neural network model to obtain deep local features. These features are then input into a fully connected neural network layer used to predict the target data, outputting the spatiotemporal relationship between the predictor sequence and the target data. Specifically:

[0075] The extracted temporal features are fed into a convolutional neural network model. The convolutional and pooling layers in the two-layer convolutional neural network extract deep local features of the temporal features. The feature matrix is ​​then output through a fully connected layer to obtain the spatiotemporal relationship between the predictor sequence and the target data.

[0076] In this embodiment, the parameters of the GRU-CNN model are set as follows: learning rate 0.001, number of iterations 250, number of GRU hidden nodes 84, number of CNN convolutional kernels 25, CNN convolutional kernel size 7, batch size 128, activation function ReLU, loss function and fitness function CCE, and regularization optimization 0.1.

[0077] S4. Obtain historical data and time scales of standardized precipitation evapotranspiration drought index at relevant time scales, capture influencing factors in the time series data to extract time features, integrate these features into the Prophet model for prediction, and extract time features from the historical standardized precipitation evapotranspiration data. The specific content is as follows:

[0078] y(t) = g(t) + s(t) + h(t) + ε(t)

[0079] Where g(t) represents the trend of the time series in the non-periodic region; s(t) represents the periodic term; h(t) represents the holiday term; and ε(t) represents the residual term.

[0080] Influencing factors include trends, cyclical patterns, and the impact of special holidays.

[0081] In this embodiment, the parameters are set as follows: changepoint_prior_scale is 0.05; interval_width is 0.8; growth is linear; mcmc_samples is 0; changepoint_range is 0.9; yearly_seasonality is True; seasonality_mode is additive; the model learning rate is 0.001, which maintains the model's convergence speed without causing divergence; the number of iterations is 150; the activation function is ReLU; and MAE is the model loss function. After repeated experiments and comparisons, the above parameter values ​​are the optimal parameters.

[0082] S5, such as Figure 4 As shown, the dataset in step S2 is divided into a training set and a test set in an 8:2 ratio to evaluate the performance of the long-term drought index prediction model. The specific details are as follows:

[0083] The prediction performance of the Prophet-GRU-CNN model is evaluated using mean absolute error, root mean square error, Nash efficiency coefficient, and mean absolute percentage error. The specific formula is as follows:

[0084]

[0085]

[0086]

[0087]

[0088] Where MAE represents mean absolute error, RMSE represents root mean square error, NSE represents Nash efficiency coefficient, T represents time, MAPE represents mean absolute percentage error, m represents the number of data points, and x represents the mean absolute percentage error. i Represents the observed value. Indicates the predicted value; Represents the observed value at time t. Represents the simulated value at time t. This represents the average value of the observed values.

[0089] Figure 4 (a) represents the predicted indicators of the SPEI01 drought index for the Zhenxiong site from the five models. Figure 4 (b) represents the predicted indicators of the SPEI03 drought index for the Zhenxiong site from the five models. Figure 4 (c) represents the prediction indicators of the SPEI06 drought index for the Zhenxiong site by the five models. Figure 4 (d) represents the predicted indicators of the five models for the SPEI012 drought index at the Zhenxiong site. It can be seen that the Prophet-GRU-CNN model outperforms the other four comparative models in all indicators. Specifically, the Prophet-GRU-CNN model outperforms the GRU-CNN model, Prophet model, GRU model, and CNN model by an average of 31.07%, 56.79%, 70.27%, and 85.64% in MAE, 1.55%, 4.36%, 8.21%, and 15.84% in NSE, 43.45%, 61.84%, 71.98%, and 79.44% in RMSE, and 22.06%, 53.21%, 71.43%, and 81.38% in MAPE.

[0090] S6. The outputs of the GRU-CNN model and the Prophet model are concatenated to create a combined matrix. This matrix contains the predictions from both models. The combined matrix is ​​then fed into an additional GRU model to further integrate the information from both models, improving the prediction performance of the drought index. Two fully connected layers are added after the GRU model to learn higher-level features and patterns, transforming the input into the final prediction result. The test samples are then fed into the Prophet-GRU-CNN model to obtain the standardized precipitation evapotranspiration drought index prediction results.

[0091] S7. To verify the predictive ability of the long-term drought index output by the model, this invention selects the fitting effect of the model output results to the real data at a certain site in Zhenxiong as a case study for the prediction of the SPEI drought index. The superiority of the algorithm is compared between the prediction performance of the Prophet-GRU-CNN model and other comparative models at this site. The prediction sequences of the five models on the SPEI index at different time scales are compared as follows: Figure 5 As shown. From Figure 5 (a) to Figure 5 In (d), among the prediction effects of the five models on the SPEI drought index at various time scales, it can be seen that the Prophet-GRU-CNN model has the best fit to the real data in terms of the time series curves predicted, while the CNN model has the worst fit to the real data. The prediction effects of the different models vary. As the time scale of the SPEI drought index increases, the frequency of change of each index curve gradually decreases.

[0092] The 35 Yunnan Province sites selected in this invention belong to different regions with significant climate differences. The Prophet-GRU-CNN model outperformed other comparative models on all 35 Yunnan Province sites, which proves that the model has strong adaptability to data and universality in time series prediction.

[0093] In summary, this invention discloses a long-term drought index prediction method based on Prophet-GRU-CNN. This method constructs a long-term drought index prediction model based on the Prophet-GRU-CNN deep learning model, and includes the following steps: correlation analysis between meteorological factor data and the SPEI drought index at various scales to select suitable predictors; inputting the time series of the selected predictors into the GRU module to extract temporal features; inputting the extracted temporal features into the CNN model to extract deep local features, obtaining the spatiotemporal relationship between the predictor sequence and the target data; inputting historical SPEI drought index data and the time scale of the data at relevant time scales into the Prophet module; concatenating the outputs of the CNN model and the Prophet model, inputting them into the GRU model, and then processing them through two fully connected layers to output the prediction result. The Prophet-GRU-CNN model proposed in this invention solves the problems of insufficient data volume, deep feature extraction, and combination of multiple algorithms. While supplementing the amount of data, it enhances the absorption of temporal and spatial features of the data, improves the utilization rate of historical data, and can also mine the deep mapping between predictor and label value. The combination of the three greatly improves the prediction effect of drought index on long time scale.

[0094] This invention also proposes a long-term drought index prediction system based on Prophet-GRU-CNN, including a predictor acquisition module, a temporal feature extraction module, a module for acquiring the spatiotemporal relationship between the predictor sequence and target data, a module for extracting the temporal features of historical data, a model evaluation module, and a computer program that can run on a processor. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention embodiment, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention embodiment.

[0095] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0096] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0097] The above description is merely an exemplary embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A long-term drought index prediction method based on Prophet-GRU-CNN, comprising: S1. Conduct correlation analysis between meteorological factor data and standardized precipitation evapotranspiration drought index at various scales to obtain predictive factors whose correlation with standardized precipitation evapotranspiration drought index is greater than a set threshold. S2. Based on the predictive factors in step S1, select time series within a set time period, preprocess them to form a dataset; select the standardized precipitation evapotranspiration drought index as the target label value, input the predictive factor time series into the gated recurrent unit model, and extract time features. S3. After the gated recurrent unit model, add a convolutional neural network layer, input the extracted temporal features into the convolutional neural network model to obtain deep local features, input the features into the fully connected neural network layer, and output the spatiotemporal relationship between the predictor sequence and the target data. S4. Obtain historical data and time scale of standardized precipitation evapotranspiration drought index at relevant time scales, capture influencing factors in time series data to extract time features, integrate these features into the Prophet model for prediction, and extract time features from standardized precipitation evapotranspiration historical data. S5. Based on steps S2-S4, obtain the Prophet-GRU-CNN model. Divide the dataset from step S2 into a training set and a test set according to a preset ratio, and evaluate the performance of the Prophet-GRU-CNN model; specifically: The outputs of the GRU-CNN model and the Prophet model are concatenated to create a combined matrix. This matrix contains the predictions of both models. The combined matrix is ​​then input into an additional GRU model to further integrate the information from both models. Two fully connected layers are added after the GRU model to learn higher-level features and patterns, and the input is transformed into the final prediction result. S6. Input the sample to be tested into the Prophet-GRU-CNN model to obtain the standardized precipitation evapotranspiration drought index prediction results.

2. The long-term drought index prediction method based on Prophet-GRU-CNN according to claim 1, characterized in that, In step S1, the predictor factors include the following: Predictive factors were selected, including precipitation, 10m wind speed, relative humidity, peak logarithmic solar radiation, evaporation, runoff, convective potential energy, and net solar radiation. Pearson correlation analysis was used to analyze the standardized precipitation evapotranspiration drought index. The correlation coefficient between each meteorological data point and the standardized precipitation evapotranspiration drought index at each station was measured. The specific formula is as follows: ； Where, x i and y i This represents the i-th pair of observations in the sample data. This represents the average value of x. This represents the average value of y. This represents the correlation coefficient.

3. The long-term drought index prediction method based on Prophet-GRU-CNN according to claim 1, characterized in that, Step S2, extracting time features, includes the following sub-steps: S201. Clean the time series data, impute missing values, and integrate the data, and then normalize the data to obtain a dataset of the corresponding dimensions. S202. Using the gated loop unit model, time features are extracted by controlling the update and reset gates. The specific formula is as follows: ； ； ； ； in, This indicates a reset of the gate output. This indicates an update to the gate output. This represents the candidate state at the current moment. This indicates the final output. This represents the hidden state at time t-1. This represents the input at time t. This represents the sigmoid activation function. and They represent x respectively t and and The weighting relationship between them and They represent h respectively t-1 and and The weighting relationship between them express The corresponding bias term, express The corresponding bias term, and They represent and and The weighting relationship between them express The corresponding bias term, This represents the element-wise product of vectors.

4. The long-term drought index prediction method based on Prophet-GRU-CNN according to claim 1, characterized in that, In step S3, the spatiotemporal relationship between the output predictor sequence and the target data includes the following: The extracted temporal features are fed into a convolutional neural network model. The convolutional and pooling layers in the two-layer convolutional neural network extract deep local features of the temporal features. The feature matrix is ​​then output through a fully connected layer to obtain the spatiotemporal relationship between the predictor sequence and the target data.

5. The long-term drought index prediction method based on Prophet-GRU-CNN according to claim 1, characterized in that, In step S4, the specific formula for extracting the temporal features from the standardized historical precipitation evapotranspiration data is as follows: ； in, This indicates the trend of time series changes in a non-periodic context; Indicates a periodic term; Indicates holiday items; Indicates the remaining terms; Influencing factors include trends, cyclical patterns, and the impact of special holidays.

6. The long-term drought index prediction method based on Prophet-GRU-CNN according to claim 1, characterized in that, In step S5, the prediction performance of the Prophet-GRU-CNN model is tested using mean absolute error, root mean square error, Nash efficiency coefficient, and mean absolute percentage error. The specific formula is as follows: ； ； ； ； Where MAE represents mean absolute error, RMSE represents root mean square error, NSE represents Nash efficiency coefficient, T represents time, MAPE represents mean absolute percentage error, and m represents the number of data points. Represents the observed value. Indicates the predicted value; Represents the observed value at time t. Represents the simulated value at time t. This represents the average value of the observed values.

7. A long-term drought index prediction system based on Prophet-GRU-CNN, characterized in that, include The predictor acquisition module is used to perform correlation analysis between meteorological factor data and standardized precipitation evapotranspiration drought index at various scales, and to obtain predictors that have a correlation with the standardized precipitation evapotranspiration drought index greater than a set threshold. The time feature extraction module is used to select time series within a set time period based on the predictor factors in the predictor acquisition module, preprocess them to form a dataset; select the standardized precipitation evapotranspiration drought index as the target label value, input the predictor time series into the gated recurrent unit model, and extract time features; The module for obtaining the spatiotemporal relationship between the predictor sequence and the target data is used to input the extracted temporal features into the convolutional neural network model, obtain deep local features, and output the spatiotemporal relationship between the predictor sequence and the target data. The historical data time feature extraction module is used to obtain historical data of standardized precipitation evapotranspiration drought index and the time scale of the data at relevant time scales, capture the influencing factors in time series data to extract time features, integrate the features into the Prophet model for prediction, and extract the time features in the standardized precipitation evapotranspiration historical data. Specifically: The outputs of the GRU-CNN model and the Prophet model are concatenated to create a combined matrix. This matrix contains the predictions of both models. The combined matrix is ​​then input into an additional GRU model to further integrate the information from both models. Two fully connected layers are added after the GRU model to learn higher-level features and patterns, and the input is transformed into the final prediction result. The model evaluation module is used to divide the dataset in the time feature extraction module into training and test sets according to a preset ratio to evaluate the performance of the long-term drought index prediction model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the method of any one of claims 1 to 6.

Images