CEEMDAN-GRU-based short-time precipitation prediction method
By preprocessing and decomposing hourly precipitation data and combining with GRU model, the lag problem of traditional models when processing long-term memory data is solved, achieving more efficient short-term precipitation prediction, significantly improving prediction accuracy and timeliness .
Patent Information
- Application Number
- CN202411340481.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2025-01-07
AI Technical Summary
Traditional rainfall prediction models are difficult to process input data from long-term memory, resulting in the impact of prediction accuracy and lag, making it difficult to achieve timely and accurate rainfall prediction.
Using the short-term precipitation prediction method based on CEEMDAN-GRU, the hourly precipitation data is preprocessed, and decomposed into sub-sequences and residual sequences in different frequency domains, and input them into the GRU model for prediction, eliminating the time series prediction , improve prediction timeliness and accuracy.
It effectively eliminates the lag of traditional models, improves the timeliness and accuracy of precipitation prediction, significantly improves the prediction accuracy, and can more accurately predict short-term precipitation in the future.
Smart Images
Figure CN119270387A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the technical field of meteorological forecasting, and in particular, relates to a short-term precipitation forecasting method based on CEEMDAN-GRU. Background Art
[0002] Precipitation forecasting plays an irreplaceable role in urban transportation, agricultural systems, water resources planning, and flood warning. In the context of global climate change, the frequency of extreme precipitation events and extreme drought events is on the rise, which has brought many uncertain challenges to agricultural production, ecological environment, and economic and social sustainable development. Therefore, it is very important to obtain rainfall information more quickly and accurately. Since the climate system is affected by many factors, although precipitation has certain regularity, its complexity and uncertainty are strong, which makes precipitation forecasting an extremely difficult task.
[0003] Traditional rainfall forecasting is mainly based on satellite cloud images and weather radar detection data. Satellite cloud images can fully reflect cloud information without geographical restrictions and can predict large-scale weather changes. Nizar et al. obtained cloud top temperature and cloud particle effective radius from satellite images, extracted characteristic values, and used logistic regression to analyze the probability relationship between these characteristic values and extreme rainfall events, thereby predicting extreme rainfall more than 6 hours in advance.
[0004] With the accumulation of data and the development of artificial intelligence, more and more machine learning algorithms are used to predict rainfall to reduce the errors caused by traditional observation methods, and machine learning methods have been widely used in precipitation forecasting. However, traditional machine learning methods have an obvious limitation, that is, it is difficult to process long-term memory input data, which has an adverse effect on prediction accuracy. Summary of the invention
[0005] In view of this, the purpose of this application is to provide a short-term precipitation prediction method based on CEEMDAN-GRU, which effectively eliminates the lag of traditional models in time series prediction, improves the timeliness and accuracy of the prediction, and significantly improves the prediction accuracy of precipitation.
[0006] The present application provides a short-term precipitation prediction method based on CEEMDAN-GRU, the method comprising:
[0007] S1. Preprocessing the hourly precipitation data in a specific time period in the past to obtain a time series signal that can characterize the precipitation;
[0008] S2. Decomposing the time series signal into a plurality of subsequences and residual sequences in different frequency domains by using the CEEMDAN method;
[0009] S3. Input multiple subsequences and residual sequences into the GRU model in sequence, and add the prediction results of each sequence to obtain the prediction result of future short-term precipitation.
[0010] Furthermore, the hourly precipitation data in a specific time period in the past is preprocessed to obtain a time series signal capable of representing the precipitation, including:
[0011] S1a, obtaining meteorological data of a target area in a specific time period in the past to obtain hourly precipitation data in the specific time period in the past;
[0012] S1b. The hourly precipitation data is sequentially supplemented with missing values, removed with outliers, and normalized to obtain a time series signal capable of representing precipitation at consecutive moments.
[0013] Furthermore, the missing value supplementation, outlier removal and normalization processing are performed on the hourly precipitation data in sequence, including:
[0014] For the missing data value in the hourly precipitation data, if the precipitation within 3 hours before and after the missing data value is 0, the missing data value is set to 0;
[0015] If the precipitation in any time period within 3 hours before and after the missing data value is not 0, the missing data value is set to the average precipitation within 3 hours before and after;
[0016] For an abnormal data value in the hourly precipitation data, if the precipitation within 3 hours before and after the abnormal data value is 0, the abnormal data value is set to 0;
[0017] If the precipitation in any time period within 3 hours before and after the abnormal data value is not 0, the abnormal data value is set to the average precipitation within 3 hours before and after;
[0018] The hourly precipitation data after missing value supplementation and outlier removal are normalized to eliminate the influence of data dimension and obtain a time series signal that can characterize precipitation at continuous moments.
[0019] Furthermore, the CEEMDAN method is used to decompose the time series signal into multiple subsequences and residual sequences in different frequency domains, including:
[0020] S2a, adding N groups of Gaussian white noise to the time series signal to obtain a new time series signal;
[0021] S2b, perform CEEMDAN decomposition on the new time series signal by the following formula to obtain the first subsequence:
[0022]
[0023] In the formula, is the IMF1 component obtained by CEEMDAN decomposition, m is the number of times Gaussian white noise is added m=1, 2, ···, N, E(k) is the Kth IMF component obtained by EMD decomposition, X(t) is the time series signal, ε 0 is the noise factor, V m Gaussian white noise that satisfies the standard normal distribution;
[0024] S2c, calculate the first-order residual signal using the following formula:
[0025]
[0026] S2d, adding pairs of positive and negative Gaussian white noise to the first-order residual signal, and performing CEEMDAN decomposition using the following formula to obtain a second subsequence:
[0027]
[0028] In the formula, is the IMF2 component decomposed by CEEMDAN, is the modal component obtained by EMD decomposition;
[0029] S2e, calculate the second-order residual signal using the following formula:
[0030]
[0031] S2f, repeating steps S2d-S2e until a monotone residual signal is calculated to determine a plurality of subsequences in different frequency domains, and determining the monotone residual signal as a residual sequence.
[0032] Furthermore, the GRU model includes: an input layer, a single-layer neural network including 50 GRU units, and a fully connected layer.
[0033] Furthermore, the GRU model is trained in the following manner:
[0034] Obtain historical meteorological data of the target area to obtain historical hourly precipitation data;
[0035] The historical hourly precipitation data is sequentially supplemented with missing values, removed with outliers, and normalized to obtain a historical time series signal capable of characterizing precipitation at consecutive moments;
[0036] Performing CEEMDAN decomposition on the historical time series signal to obtain multiple subsequences and residual sequences of the historical time series signal;
[0037] The GRU model is trained using multiple subsequences and residual sequences of the historical time series signal.
[0038] The short-term precipitation prediction method based on CEEMDAN-GRU provided in this application preprocesses the hourly precipitation data in a specific time period in the past to obtain a time series signal that can characterize the precipitation; using the CEEMDAN method, the time series signal is decomposed into multiple subsequences and residual sequences in different frequency domains; the multiple subsequences and residual sequences are sequentially input into the GRU model, and the prediction results of each sequence are added to obtain the prediction result of the future short-term precipitation. The method effectively eliminates the lag of the traditional model in time series prediction, improves the timeliness and accuracy of the prediction, and significantly improves the prediction accuracy of precipitation. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 A flow chart of a short-term precipitation prediction method based on CEEMDAN-GRU provided in an embodiment of the present application is shown;
[0040] Figure 2 A schematic diagram showing multiple subsequences and residual sequences in different frequency domains obtained by CEEMDAN decomposition provided in an embodiment of the present application;
[0041] Figure 3 A comparison chart of the precipitation test value and the actual value in Shenyang area in the next hour provided by the embodiment of the present application is shown;
[0042] Figure 4 A comparison chart of the precipitation test value and the actual value in the Liaozhong area in the next hour provided by the embodiment of the present application is shown;
[0043] Figure 5 A comparison chart of the precipitation test value and the actual value in the next hour in Kangping area provided in an embodiment of the present application is shown. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical solution and advantages of the technical solution more clear, the technical solution is further described in detail below in conjunction with specific implementation methods. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the technical solution.
[0045] First, the CEEMDAN algorithm used in this application is described:
[0046] The CEEMDAN algorithm is an improvement on the EMD (Empirical Mode Decomposition) and EEMD (Ensemble Empirical Mode Decomposition) algorithms. The algorithm can adaptively decompose the original data into multiple intrinsic mode functions (IMFs) and residuals of different frequencies and lengths, and is suitable for processing nonlinear and non-stationary time series data. During the decomposition process, the CEEMDAN algorithm adaptively adjusts the noise coefficient and introduces Gaussian noise with different signal-to-noise ratios into the signal to be decomposed. This process aims to avoid modal aliasing problems and eliminate the interference of false information at the same time.
[0047] Compared with algorithms such as EMD and EEMD, the CEEMDAN algorithm has the following advantages: 1) By introducing an additional signal-to-noise ratio, the noise level in each decomposition process is effectively controlled; 2) The signal is decomposed into multiple independent mode functions (IMFs), which can completely reconstruct the original data with almost no noise components; 3) The number of tests is less and the decomposition efficiency is higher.
[0048] Next, the technical solution of this application is explained:
[0049] See also Figure 1 , Figure 1 FIG. 8 is a flow chart of a short-term precipitation prediction method based on CEEMDAN-GRU provided in an embodiment of the present application. Figure 1 As shown, the method includes:
[0050] S101. Preprocessing the hourly precipitation data within a specific time period in the past to obtain a time series signal capable of representing the precipitation.
[0051] In this step, the past specific time period refers to the past N hours up to the current moment, where N is set after analyzing the historical precipitation data. If the analysis results show that the precipitation in a certain hour is related to the precipitation in the previous 10 hours, N is set to 10. Of course, N can also be set to other values according to the specific situation, and this application does not make any limitation here.
[0052] In specific implementation, the time series signal that can characterize precipitation can be obtained by the following methods:
[0053] Step 1011: Acquire meteorological data of the target area in a specific time period in the past to obtain hourly precipitation data in the specific time period in the past.
[0054] In this step, meteorological data of the target area is obtained, and the precipitation data of the previous hour is recorded at a time interval of 1 hour each time to obtain the hourly precipitation data in the past specific time period.
[0055] As an example, when the specific time period is set to 10, precipitation data is continuously recorded 10 times at a time interval of 1 hour, that is, the hourly precipitation data in the past specific time period is obtained.
[0056] Step 1012: supplement missing values, remove outliers, and normalize the hourly precipitation data in sequence to obtain a time series signal that can characterize precipitation at consecutive moments.
[0057] Specifically, for a missing data value in the hourly precipitation data, if the precipitation within 3 hours before and after the missing data value is 0, the missing data value is set to 0;
[0058] If the precipitation in any time period within 3 hours before and after the missing data value is not 0, the missing data value is set to the average precipitation within 3 hours before and after;
[0059] For an abnormal data value in the hourly precipitation data, if the precipitation within 3 hours before and after the abnormal data value is 0, the abnormal data value is set to 0;
[0060] If the precipitation in any time period within 3 hours before and after the abnormal data value is not 0, the abnormal data value is set to the average precipitation within 3 hours before and after;
[0061] The hourly precipitation data after missing value supplementation and outlier removal are normalized to eliminate the influence of data dimension and obtain a time series signal that can characterize precipitation at continuous moments.
[0062] In this step, the MinMaxScaler function is used to normalize the hourly precipitation data after missing value supplementation and outlier removal using the following formula (1):
[0063]
[0064] Where, X min , X max They are the minimum and maximum values of hourly precipitation data respectively.
[0065] S102: Decompose the time series signal into multiple subsequences and residual sequences in different frequency domains by using the CEEMDAN method.
[0066] In specific implementation, multiple subsequences and residual sequences in different frequency domains can be obtained in the following manner:
[0067] Step 1021: Add N groups of Gaussian white noise to the time series signal to obtain a new time series signal.
[0068] Step 1022: Perform CEEMDAN decomposition on the new time series signal using the following formula (2) to obtain the first subsequence:
[0069]
[0070] In the formula, is the IMF1 component obtained by CEEMDAN decomposition, m is the number of times Gaussian white noise is added m=1, 2, ···, N, E(k) is the Kth IMF component obtained by EMD decomposition, X(t) is the time series signal, ε 0 is the noise factor, V m is Gaussian white noise that satisfies the standard normal distribution.
[0071] Step 1023: Calculate the first-order residual signal using the following formula (3):
[0072]
[0073] Step 1024: Add paired positive and negative Gaussian white noise to the first-order residual signal, and perform CEEMDAN decomposition using the following formula (4) to obtain a second subsequence:
[0074]
[0075] In the formula, is the IMF2 component decomposed by CEEMDAN, are the modal components obtained by EMD decomposition.
[0076] Step 1025: Calculate the second-order residual signal using the following formula (5):
[0077]
[0078] Step 1026: Repeat steps 1024-1025 until a monotone residual signal is calculated to determine a plurality of subsequences in different frequency domains, and determine the monotone residual signal as a residual sequence.
[0079] In this step, the relationship between the time series signal X(t) and the multiple subsequences and residual sequences obtained by decomposition is also expressed as the following formula (6):
[0080]
[0081] Here, from IMF1 to IMFi, the complexity of the sequence gradually decreases.
[0082] S103, inputting multiple subsequences and residual sequences into the GRU model in sequence, and adding the prediction results of each sequence to obtain the prediction result of future short-term precipitation.
[0083] The GRU model includes: an input layer, a single-layer neural network containing 50 GRU units, and a fully connected layer.
[0084] In this step, GRU (Gated Recurrent Unit) is a type of RNN, which is used to solve the problems of long-term memory loss and gradient vanishing or exploding in back propagation in traditional RNN when dealing with practical problems. Taking advantage of the GRU model's ability to capture nonlinear dynamic changes in time series, the GRU model is combined with the CEEMDAN algorithm to improve the accuracy and stability of short-term precipitation prediction.
[0085] Here, after many experiments, the structure of the GRU model was determined to be: an input layer, a single-layer neural network containing 50 GRU units, and a fully connected layer. This model structure performs well in predicting short-term precipitation.
[0086] In specific implementation, the GRU model can be trained in the following ways:
[0087] Step 201: Acquire historical meteorological data of a target area to obtain historical hourly precipitation data.
[0088] As an example, the observation data of three meteorological observation stations including Shenyang and its surrounding Liaozhong and Kangping are obtained from the Liaoning Provincial Meteorological Bureau. The three meteorological observation stations have a wide coverage and can more comprehensively reflect the changing pattern of precipitation in the region and the effectiveness of the model. The observation data covers four years of meteorological data from January 1, 2019 to December 31, 2022. The historical hourly precipitation data of the three regions are recorded at intervals of 1 hour.
[0089] Step 202: supplement missing values, remove outliers, and normalize the historical hourly precipitation data in sequence to obtain a historical time series signal that can characterize precipitation at consecutive moments.
[0090] As an example, for the historical hourly precipitation data of the same region, missing value supplementation, outlier removal and normalization are performed in sequence to obtain the historical time series signal of the region.
[0091] Step 203: Perform CEEMDAN decomposition on the historical time series signal to obtain multiple subsequences and residual sequences of the historical time series signal.
[0092] As an example, CEEMDAN decomposition is performed on a historical time series signal to obtain multiple subsequences and residual sequences of the historical time series signal.
[0093] Here, please see Figure 2 , Figure 2 The figure shows a schematic diagram of multiple subsequences and residual sequences in different frequency domains obtained by CEEMDAN decomposition provided by an embodiment of the present application. In the figure, the horizontal axis is the number of time series of precipitation, in hours (h), and the vertical axis is the value of precipitation in mm. Figure 2 There are 19 IMF components, namely 19 subsequences and one residual sequence, and the complexity of the sequences IMF1-IMF19 gradually decreases.
[0094] Step 204: Use multiple subsequences and residual sequences of the historical time series signal to train the GRU model.
[0095] As an example, set the parameters of the GRU model as follows: Optimizer is Adam, Iterations is 20, LossFunction is MSE, Window Length is 10, and Batch is 16.
[0096] In order to test the performance of the short-term precipitation prediction method based on CEEMDAN-GRU, the test values of the next hour obtained by the short-term precipitation prediction method based on CEEMDAN-GRU are compared with the true values for three regions. Figure 3 , 4 , 5 are comparison diagrams of the precipitation test value and the true value in the next hour in Shenyang, Liaozhong and Kangping areas. It can be seen from the figure that the short-term precipitation prediction method based on CEEMDAN-GRU in this application has good effectiveness and stability.
[0097] The above contents are only preferred embodiments of the present invention. For ordinary technicians in this field, many changes can be made in the specific implementation methods and application scopes based on the ideas of the technical content. As long as these changes do not deviate from the concept of the present invention, they all fall within the scope of protection of this patent.
Claims
1. A short-term precipitation prediction method based on CEEMDAN-GRU, characterized in that: The method comprises: S1. Preprocessing the hourly precipitation data in a specific time period in the past to obtain a time series signal that can characterize the precipitation; S2. Decomposing the time series signal into a plurality of subsequences and residual sequences in different frequency domains by using the CEEMDAN method; S3. Input multiple subsequences and residual sequences into the GRU model in sequence, and add the prediction results of each sequence to obtain the prediction result of future short-term precipitation.
2. The method according to claim 1, characterized in that The preprocessing of the hourly precipitation data in the past specific time period to obtain a time series signal capable of representing the precipitation includes: S1a, obtaining meteorological data of a target area in a specific time period in the past to obtain hourly precipitation data in the specific time period in the past; S1b. The hourly precipitation data is sequentially supplemented with missing values, removed with outliers, and normalized to obtain a time series signal capable of representing precipitation at consecutive moments.
3. The method according to claim 2, characterized in that The missing value supplementation, outlier removal and normalization processing of the hourly precipitation data are performed in sequence, including: For the missing data value in the hourly precipitation data, if the precipitation within 3 hours before and after the missing data value is 0, the missing data value is set to 0; If the precipitation in any time period within 3 hours before and after the missing data value is not 0, the missing data value is set to the average precipitation within 3 hours before and after; For an abnormal data value in the hourly precipitation data, if the precipitation within 3 hours before and after the abnormal data value is 0, the abnormal data value is set to 0; If the precipitation in any time period within 3 hours before and after the abnormal data value is not 0, the abnormal data value is set to the average precipitation within 3 hours before and after; The hourly precipitation data after missing value supplementation and outlier removal are normalized to eliminate the influence of data dimension and obtain a time series signal that can characterize precipitation at continuous moments.
4. The method according to claim 1, characterized in that The CEEMDAN method is used to decompose the time series signal into multiple subsequences and residual sequences in different frequency domains, including: S2a, adding N groups of Gaussian white noise to the time series signal to obtain a new time series signal; S2b, perform CEEMDAN decomposition on the new time series signal by the following formula to obtain the first subsequence: In the formula, is the IMF1 component obtained by CEEMDAN decomposition, m is the number of times Gaussian white noise is added m=1, 2, ···, N, E(k) is the Kth IMF component obtained by EMD decomposition, X(t) is the time series signal, ε0 is the noise coefficient, V m Gaussian white noise that satisfies the standard normal distribution; S2c, calculate the first-order residual signal using the following formula: S2d, adding pairs of positive and negative Gaussian white noise to the first-order residual signal, and performing CEEMDAN decomposition using the following formula to obtain a second subsequence: In the formula, is the IMF2 component decomposed by CEEMDAN, is the modal component obtained by EMD decomposition; S2e, calculate the second-order residual signal using the following formula: S2f, repeating steps S2d-S2e until a monotone residual signal is calculated to determine a plurality of subsequences in different frequency domains, and determining the monotone residual signal as a residual sequence.
5. The method according to claim 1, characterized in that The GRU model includes: an input layer, a single-layer neural network including 50 GRU units, and a fully connected layer.
6. The method according to claim 5, characterized in that The GRU model is trained in the following way: Obtain historical meteorological data of the target area to obtain historical hourly precipitation data; The historical hourly precipitation data is sequentially supplemented with missing values, removed with outliers, and normalized to obtain a historical time series signal capable of characterizing precipitation at consecutive moments; Performing CEEMDAN decomposition on the historical time series signal to obtain multiple subsequences and residual sequences of the historical time series signal; The GRU model is trained using multiple subsequences and residual sequences of the historical time series signal.