Time sequence characteristic analysis method and system based on multi-dimensional data

Abstract

The invention relates to the technical field of data processing, and discloses a time sequence feature analysis method and system based on multi-dimensional data. The method comprises the following steps: performing spatial-temporal feature extraction and dimension reduction processing on a multi-dimensional time sequence precipitation data set to obtain a precipitation spatial-temporal feature vector set; performing multi-scale spatio-temporal pattern recognition and clustering analysis on the precipitation spatio-temporal feature vector set to obtain a precipitation spatio-temporal pattern library, a clustering result and an abnormal precipitation pattern list; constructing a deep learning rainfall prediction model to obtain an end-to-end global atmospheric reanalysis rainfall fusion prediction model; performing dynamic updating and online learning on the global atmospheric reanalysis rainfall fusion prediction model to obtain a target rainfall space-time prediction model; and performing multi-dimensional space-time analysis on the target rainfall space-time prediction model to obtain a global atmosphere reanalysis rainfall fusion data analysis report and a weather processing assembly line scheduling strategy. According to the invention, the accuracy of time sequence feature analysis based on multi-dimensional data is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a method and system for analyzing time series characteristics based on multidimensional data. Background Art

[0002] In the prior art, weather observation and forecasting systems usually rely on a single data source, such as ground observation stations or satellite remote sensing data. These systems make precipitation forecasts by analyzing a single data source. However, the data from a single data source has limitations and cannot fully reflect the complex changes in the atmosphere. In addition, existing precipitation forecast models mainly rely on traditional statistical methods and numerical simulation methods, which are difficult to capture complex nonlinear spatiotemporal characteristics, resulting in low forecast accuracy. At the same time, traditional precipitation forecasting methods are difficult to dynamically update and adaptively learn, and cannot respond to weather changes in a timely manner, affecting the accuracy and timeliness of the forecast.

[0003] The deficiency of the existing technology is that there is a lack of an effective multi-source data fusion method that can make full use of the multi-dimensional time series data of weather observation stations, satellite remote sensing and ground observation stations to improve the accuracy of precipitation forecasts. In addition, the existing methods lack in-depth analysis and mining of precipitation spatiotemporal characteristics, and cannot effectively identify abnormal precipitation patterns and multi-scale spatiotemporal patterns, resulting in limited prediction capabilities for extreme weather events. In addition, the existing precipitation forecast model lacks dynamic update and online learning capabilities, cannot adapt to rapidly changing weather conditions, and is difficult to provide high-precision, real-time precipitation forecasts. Summary of the invention

[0004] The present application provides a method and system for analyzing time series features based on multidimensional data, which are used to improve the accuracy of time series feature analysis based on multidimensional data.

[0005] In a first aspect, the present application provides a method for analyzing time series features based on multidimensional data, and the method comprises: collecting and preprocessing multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations to obtain a standardized multidimensional time series precipitation data set; performing spatiotemporal feature extraction and dimensionality reduction processing on the multidimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set; performing multi-scale spatiotemporal pattern recognition and clustering analysis on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results and an abnormal precipitation pattern list; constructing a deep learning precipitation prediction model based on the precipitation spatiotemporal pattern library, the clustering results and the abnormal precipitation pattern list to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model; dynamically updating and online learning the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model; performing multidimensional spatiotemporal analysis on the target precipitation spatiotemporal prediction model to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy.

[0006] In a second aspect, the present application provides a time series feature analysis system based on multidimensional data, the time series feature analysis system based on multidimensional data comprising:

[0007] The acquisition module is used to collect and preprocess the multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations to obtain a standardized multi-dimensional time series precipitation dataset;

[0008] A dimension reduction module is used to extract spatiotemporal features and perform dimension reduction processing on the multidimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set;

[0009] A clustering module is used to perform multi-scale spatiotemporal pattern recognition and cluster analysis on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results and an abnormal precipitation pattern list;

[0010] A construction module is used to construct a deep learning precipitation prediction model based on the precipitation spatiotemporal pattern library, the clustering results and the abnormal precipitation pattern list to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model;

[0011] An updating module is used to dynamically update and online learn the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model;

[0012] The analysis module is used to perform multi-dimensional spatiotemporal analysis on the target precipitation spatiotemporal prediction model to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy.

[0013] In the technical solution provided by the present application, by collecting and preprocessing the multi-source raw data of weather observation stations, satellite remote sensing and ground observation stations, a standardized multi-dimensional time series precipitation data set can be obtained. This data set combines the advantages of multi-source data, avoids the limitations of a single data source, and makes the data more comprehensive and accurate. On this basis, the multi-dimensional time series precipitation data set is subjected to spatiotemporal feature extraction and dimensionality reduction processing to obtain a precipitation spatiotemporal feature vector set. In this way, the complex spatiotemporal characteristics of precipitation can be effectively captured, and the expression ability and processing efficiency of the data are improved. Secondly, multi-scale spatiotemporal pattern recognition and clustering analysis are performed on the precipitation spatiotemporal feature vector set, and a precipitation spatiotemporal pattern library, clustering results and a list of abnormal precipitation patterns can be constructed. This not only helps to deeply explore the spatiotemporal pattern of precipitation, but also can identify abnormal precipitation patterns and improve the prediction ability of extreme weather events. In addition, based on the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list, a deep learning precipitation prediction model is constructed to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model. This deep learning model can effectively use large-scale data for training, capture the complex nonlinear characteristics of precipitation, and significantly improve the accuracy of precipitation forecasts. Through dynamic updating and online learning, the global atmospheric reanalysis precipitation fusion prediction model is optimized to obtain the target precipitation spatiotemporal prediction model, so that the model can adaptively adjust parameters and respond to weather changes in a timely manner, thereby improving the real-time and accuracy of the forecast. Furthermore, a multi-dimensional spatiotemporal analysis is performed on the target precipitation spatiotemporal prediction model, and a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy can be generated. The precipitation forecast results can be comprehensively evaluated and analyzed. The present invention not only significantly improves the accuracy and real-time nature of precipitation forecasts, but also enhances the prediction ability of extreme weather events through the fusion of multi-source data, spatiotemporal feature extraction, multi-scale spatiotemporal pattern recognition, and construction and optimization of deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0015] Figure 1 A schematic diagram of an embodiment of a method for analyzing time series characteristics based on multidimensional data in an embodiment of the present application;

[0016] Figure 2 Schematic diagram of an embodiment of a time series feature analysis system based on multi-dimensional data in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application embodiment provides a kind of time series characteristic analysis method and system based on multidimensional data. The term "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings is used to distinguish similar objects, and need not be used to describe a specific order or precedence. It should be understood that the data used in this way can be interchangeable in appropriate circumstances, so that the embodiment described here can be implemented in a sequence other than the content illustrated or described here. In addition, the term "including" or "having" and any variation thereof are intended to cover non-exclusive inclusions, for example, the process, method, system, product or equipment comprising a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0018] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the time series feature analysis method based on multidimensional data includes:

[0019] Step S101, collecting and preprocessing multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations to obtain a standardized multi-dimensional time series precipitation data set;

[0020] Step S102, extracting spatiotemporal features and reducing dimension of the multidimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set;

[0021] Step S103, performing multi-scale spatiotemporal pattern recognition and cluster analysis on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results and an abnormal precipitation pattern list;

[0022] Step S104: construct a deep learning precipitation prediction model based on the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model;

[0023] Step S105, dynamically updating and online learning the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model;

[0024] Step S106: Perform multi-dimensional spatiotemporal analysis on the target precipitation spatiotemporal prediction model to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy.

[0025] It is understandable that the execution subject of the present application may be a time series feature analysis system based on multi-dimensional data, or may be a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0026] Specifically, the multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations are collected and preprocessed to obtain a standardized multi-dimensional time series precipitation data set. The collected data include various types of precipitation data from different times and locations, such as precipitation, precipitation intensity and precipitation frequency. These data often have different timestamps and spatial coordinates, so timestamp calibration and spatial coordinate conversion are required. Through timestamp calibration, the time of different data sources can be synchronized to ensure data consistency; through spatial coordinate conversion, the data can be converted to a unified projection system for subsequent processing. After completing time synchronization and spatial conversion, the data quality is evaluated, defective or abnormal data points are marked, and multi-source data fusion is performed. In the process of multi-source data fusion, the Kalman filter algorithm is used to eliminate noise and obtain noise reduction optimization data. The Kalman filter algorithm is an adaptive filtering algorithm that can effectively reduce the impact of noise and improve data accuracy. Then, the noise reduction optimization data is adjusted for spatiotemporal scale consistency so that the data is consistent in time and space, and the data is discretized to obtain discrete time series data. Finally, the discretized time series data are encoded with multidimensional features to form a standardized multidimensional time series precipitation dataset.

[0027] After obtaining the standardized multidimensional time series precipitation data set, spatiotemporal feature extraction and dimensionality reduction processing are performed to obtain the spatiotemporal feature vector set of precipitation. First, the fractal dimension of the multidimensional time series precipitation data set is calculated to obtain the spatiotemporal complexity characteristics. Fractal dimension is an indicator of data complexity, which can reflect the spatiotemporal complexity of precipitation data. Then, symbolic dynamics coding is performed to obtain symbolic sequence characteristics. Symbolic dynamics coding is a method of converting continuous time series into discrete symbolic sequences, which can capture the dynamic characteristics of time series. By reconstructing the phase space, the multidimensional time series precipitation data set is embedded to obtain the phase space trajectory characteristics. Phase space reconstruction is a nonlinear time series analysis method that can reveal the intrinsic structure and dynamic characteristics of time series. Then, the Lyapunov exponent is calculated for the phase space trajectory characteristics to obtain the system stability characteristics. The Lyapunov exponent is an indicator of system stability that can reflect the stability of the precipitation system. Then, recursive quantitative analysis is performed to obtain the system periodicity characteristics. Recursive quantitative analysis is a tool for analyzing the periodicity of time series, which can reveal the periodicity characteristics of precipitation data. Multiscale entropy analysis is performed to obtain the time series complexity characteristics. Multiscale entropy analysis is an indicator used to measure the complexity of time series, which can capture the multiscale complexity of precipitation data. Through the Hilbert-Huang transform, time-frequency analysis is performed on the multidimensional time series precipitation data set to obtain nonlinear time-frequency characteristics. The Hilbert-Huang transform is a nonlinear method for time-frequency analysis that can reveal the time-frequency characteristics of precipitation data. Finally, lagged mutual information analysis is performed to obtain spatiotemporal dependency characteristics. Lagged mutual information analysis is a tool for measuring the dependency between time series, which can reveal the spatiotemporal dependency of precipitation data. Nonlinear dimensionality reduction is performed on these spatiotemporal features to obtain a set of precipitation spatiotemporal feature vectors.

[0028] After obtaining the precipitation spatiotemporal feature vector set, multi-scale spatiotemporal pattern recognition and cluster analysis are performed to obtain the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list. First, the precipitation spatiotemporal feature vector set is segmented by spatiotemporal sliding window to obtain a multi-scale feature segment set. Spatiotemporal sliding window segmentation is a method for segmenting spatiotemporal data, which can capture the multi-scale characteristics of precipitation data. Next, spatiotemporal pattern encoding is performed to obtain a pattern encoding sequence. Spatiotemporal pattern encoding is a method for converting spatiotemporal features into discrete patterns, which can capture the spatiotemporal patterns of precipitation data. Then, frequent pattern mining is performed to obtain a candidate spatiotemporal pattern set. Frequent pattern mining is a tool for discovering frequently occurring patterns in data, which can reveal common patterns of precipitation data. Next, spatiotemporal association rule analysis is performed to obtain spatiotemporal association patterns. Spatiotemporal association rule analysis is a method for discovering spatiotemporal association patterns in data, which can reveal the spatiotemporal correlation of precipitation data. The spatiotemporal association patterns are clustered by the spectral clustering algorithm to obtain preliminary clustering results. The spectral clustering algorithm is a clustering method based on graph theory, which can effectively cluster spatiotemporal data. Then, fuzzy membership calculation is performed to obtain soft cluster assignment. Fuzzy membership calculation is a method used to measure the degree to which data points belong to different clusters, which can reveal the fuzzy clustering relationship of precipitation data. Next, density peak detection is performed to obtain cluster centers. Density peak detection is a density-based clustering method that can find cluster centers in data. Based on the cluster centers, the precipitation spatiotemporal feature vector set is redistributed to obtain clustering results. The clustering result is an improved clustering result that can better reveal the clustering structure of precipitation data. Then, outlier detection is performed to obtain a candidate set of abnormal precipitation patterns. Outlier detection is a method used to discover abnormal patterns in data, which can reveal abnormal patterns in precipitation data. Through the isolation forest algorithm, the abnormal precipitation pattern candidate set is anomaly quantified to obtain a list of abnormal precipitation patterns. The isolation forest algorithm is a machine learning algorithm for anomaly detection, which can effectively measure the degree of abnormality of data. Finally, the spatiotemporal correlation patterns, clustering results and abnormal precipitation pattern lists are integrated to obtain the precipitation spatiotemporal pattern library.

[0029] Based on the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list, a deep learning precipitation prediction model is constructed to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model. First, the precipitation spatiotemporal pattern library is pattern embedded to obtain a pattern embedding matrix. Pattern embedding is a method of converting patterns into vector representations, which can capture the pattern characteristics of precipitation data. Then, the clustering results are feature mapped to obtain a clustering feature tensor. Feature mapping is a method of converting clustering results into feature representations, which can capture the clustering characteristics of precipitation data. Then, the abnormal precipitation pattern list is encoded with abnormal patterns to obtain an abnormal feature vector. Abnormal pattern encoding is a method of converting abnormal patterns into vector representations, which can capture the abnormal characteristics of precipitation data. Then, the pattern embedding matrix, clustering feature tensor and abnormal feature vector are multimodally fused to obtain fused feature representation data. Multimodal feature fusion is a method of combining multiple feature representations to capture the comprehensive characteristics of precipitation data. Through the graph convolutional network, the fused feature representation data is modeled with spatiotemporal dependencies to obtain a spatiotemporal relationship graph. The graph convolutional network is a deep learning algorithm based on graph structure, which can effectively model the spatiotemporal dependency of data. Then, the attention mechanism is calculated on the spatiotemporal relationship graph to obtain weighted spatiotemporal features. The attention mechanism is a method for weighting the importance of features, which can improve the feature representation ability of precipitation data. Through the bidirectional long short-term memory network, the weighted spatiotemporal features are sequence modeled to obtain the time series prediction results. The bidirectional long short-term memory network is a deep learning algorithm for processing sequence data, which can capture the time series characteristics of precipitation data. Finally, the time series prediction results are deconvolved to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model. The deconvolution operation is a method for restoring the spatial structure of data, which can improve the accuracy of precipitation prediction.

[0030] The global atmospheric reanalysis precipitation fusion prediction model is dynamically updated and learned online to obtain the target precipitation spatiotemporal prediction model. First, the real-time precipitation observation data is segmented to obtain time series data segments. Data stream segmentation is a method of segmenting continuous data into several segments, which can capture the time series changes of precipitation data. Then, concept drift detection is performed to obtain data distribution change indicators. Concept drift detection is a method for discovering data distribution changes, which can reveal the change characteristics of precipitation data. Based on the data distribution change indicators, the global atmospheric reanalysis precipitation fusion prediction model is analyzed for parameter sensitivity to obtain a sensitive parameter set. Parameter sensitivity analysis is a method for evaluating the importance of model parameters, which can improve the adaptive ability of precipitation prediction models. Then, the sensitive parameter set is incrementally learned to obtain an updated parameter vector. Incremental learning is a method of gradually updating the model, which can improve the real-time performance of the precipitation prediction model. Then, the updated parameter vector is fused with the global atmospheric reanalysis precipitation fusion prediction model to obtain a temporary prediction model. Model fusion is a method of combining multiple model results to improve the accuracy of precipitation prediction. Then, the temporary prediction model is compressed to obtain a lightweight prediction model. Model compression is a method to reduce model complexity and improve the computational efficiency of precipitation prediction. Through federated learning, the lightweight prediction model is distributedly trained to obtain a multi-source collaborative optimization model. Federated learning is a distributed training method that can improve the collaborative optimization capability of the precipitation prediction model. Next, the multi-source collaborative optimization model is model distilled to obtain a knowledge transfer model. Model distillation is a method to transfer knowledge from a large model to a small model, which can improve the lightweight degree of the precipitation prediction model. Then, the knowledge transfer model is trained with adversarial samples to obtain a robustness enhancement model. Adversarial sample training is a method to improve the robustness of the model and enhance the anti-interference ability of the precipitation prediction model. Finally, multi-task learning is performed on the robustness enhancement model to obtain the target precipitation spatiotemporal prediction model. Multi-task learning is a method to optimize multiple tasks simultaneously, which can improve the comprehensive performance of the precipitation prediction model.

[0031] A multidimensional spatiotemporal analysis is performed on the target precipitation spatiotemporal prediction model to obtain the global atmospheric reanalysis precipitation fusion data analysis report and the weather processing pipeline scheduling strategy. First, the output of the target precipitation spatiotemporal prediction model is spatiotemporally decomposed to obtain the multi-scale precipitation prediction results. Spatiotemporal decomposition is a method of decomposing data into different scales, which can capture the multi-scale characteristics of precipitation data. Then, a spatiotemporal autocorrelation analysis is performed to obtain the evolution characteristics of precipitation patterns. Spatiotemporal autocorrelation analysis is a method for analyzing the spatiotemporal correlation of data, which can reveal the evolution characteristics of precipitation data. Then, a topological data analysis is performed to obtain the topological structure of precipitation patterns. Topological data analysis is a method for discovering the topological structure of data, which can reveal the pattern structure of precipitation data. Based on the topological structure of precipitation patterns, extreme event identification is performed to obtain extreme precipitation risk assessment. Extreme event identification is a method for discovering extreme events, which can assess the risk of precipitation data. Then, scenario simulation is performed to obtain multi-scenario precipitation impact prediction. Scenario simulation is a method for simulating different scenarios, which can predict the impact of precipitation data. Through complex network analysis, the system vulnerability assessment of multi-scenario precipitation impact forecast is carried out to obtain key influencing factors. Complex network analysis is an analysis method based on network structure, which can assess the vulnerability of precipitation data. Then, causal inference is performed to obtain the precipitation system dynamics model. Causal inference is a method for discovering causal relationships, which can reveal the dynamic mechanism of precipitation data. Based on the precipitation system dynamics model, a spatiotemporal knowledge graph is constructed to obtain a precipitation knowledge network. Spatiotemporal knowledge graph construction is a method of integrating spatiotemporal information to form a knowledge network of precipitation data. Finally, an interpretability analysis is performed to obtain a global atmospheric reanalysis precipitation fusion data analysis report. Interpretability analysis is a method for explaining model results, which can improve the interpretability of precipitation forecasts. Based on the global atmospheric reanalysis precipitation fusion data analysis report, resource optimization configuration is performed to obtain the weather processing pipeline scheduling strategy. Resource optimization configuration is a method for optimizing resource allocation, which can improve the application effect of precipitation forecasts.

[0032] For example, in a practical application, multi-source original precipitation data of a certain area are collected, including ground observation data of weather observation stations, satellite remote sensing data and ground automatic weather station data. These data have different collection frequencies in time and different distribution ranges in space. Through timestamp calibration and spatial coordinate conversion, these data are synchronized to the same time scale and space scale to form a unified multi-dimensional time series data. Then, the noise in the data is eliminated by the Kalman filter algorithm to obtain noise reduction optimization data. Then, the noise reduction optimization data is adjusted and discretized for time and space scale consistency to form a standardized multi-dimensional time series precipitation data set. Through fractal dimension calculation, symbolic dynamics coding, phase space reconstruction, Lyapunov index calculation, recursive quantitative analysis, multi-scale entropy analysis, Hilbert-Huang transform and lag mutual information analysis, the spatiotemporal complexity characteristics, symbol sequence characteristics, phase space trajectory characteristics, system stability characteristics, system periodicity characteristics, time series complexity characteristics, nonlinear time-frequency characteristics and spatiotemporal dependence characteristics of precipitation data are extracted. These features are subjected to nonlinear dimensionality reduction to obtain a precipitation spatiotemporal feature vector set. Through spatiotemporal sliding window segmentation, spatiotemporal pattern encoding, frequent pattern mining, spatiotemporal association rule analysis, spectral clustering algorithm, fuzzy membership calculation, density peak detection, outlier detection and isolation forest algorithm, multi-scale spatiotemporal pattern recognition and clustering analysis are performed on the precipitation spatiotemporal feature vector set to form a precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list. Based on the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list, a deep learning precipitation prediction model is constructed. Through pattern embedding, feature mapping, abnormal pattern encoding, multimodal feature fusion, graph convolutional network, attention mechanism, bidirectional long short-term memory network and deconvolution operation, an end-to-end global atmospheric reanalysis precipitation fusion prediction model is obtained. The model is dynamically updated and learned online, and the target precipitation spatiotemporal prediction model is formed through data stream segmentation, concept drift detection, parameter sensitivity analysis, incremental learning, model fusion, model compression, federated learning, model distillation, adversarial sample training and multi-task learning. Finally, through spatiotemporal decomposition, spatiotemporal autocorrelation analysis, topological data analysis, extreme event identification, scenario simulation, complex network analysis, causal inference, spatiotemporal knowledge graph construction and interpretability analysis, the target precipitation spatiotemporal prediction model is subjected to multidimensional spatiotemporal analysis to form a global atmospheric reanalysis precipitation fusion data analysis report, and based on this report, resource allocation is optimized and a weather processing pipeline scheduling strategy is formulated. In practical applications, the above methods can be used to make high-precision predictions of future precipitation in a certain area, identify potential extreme precipitation events, and provide optimized resource scheduling strategies to effectively deal with the impact of precipitation.

[0033] In the embodiment of the present application, by collecting and preprocessing the multi-source raw data of weather observation stations, satellite remote sensing and ground observation stations, a standardized multi-dimensional time series precipitation data set can be obtained. This data set combines the advantages of multi-source data, avoids the limitations of a single data source, and makes the data more comprehensive and accurate. On this basis, the multi-dimensional time series precipitation data set is subjected to spatiotemporal feature extraction and dimensionality reduction processing to obtain a precipitation spatiotemporal feature vector set. In this way, the complex spatiotemporal characteristics of precipitation can be effectively captured, and the expression ability and processing efficiency of the data are improved. Secondly, multi-scale spatiotemporal pattern recognition and cluster analysis are performed on the precipitation spatiotemporal feature vector set, and a precipitation spatiotemporal pattern library, clustering results and a list of abnormal precipitation patterns can be constructed. This not only helps to deeply explore the spatiotemporal pattern of precipitation, but also can identify abnormal precipitation patterns and improve the prediction ability of extreme weather events. In addition, based on the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list, a deep learning precipitation prediction model is constructed to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model. This deep learning model can effectively use large-scale data for training, capture the complex nonlinear characteristics of precipitation, and significantly improve the accuracy of precipitation forecasts. Through dynamic updating and online learning, the global atmospheric reanalysis precipitation fusion prediction model is optimized to obtain the target precipitation spatiotemporal prediction model, so that the model can adaptively adjust parameters and respond to weather changes in a timely manner, thereby improving the real-time and accuracy of the forecast. Furthermore, a multi-dimensional spatiotemporal analysis is performed on the target precipitation spatiotemporal prediction model, and a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy can be generated. The precipitation forecast results can be comprehensively evaluated and analyzed. The present invention not only significantly improves the accuracy and real-time nature of precipitation forecasts, but also enhances the prediction ability of extreme weather events through the fusion of multi-source data, spatiotemporal feature extraction, multi-scale spatiotemporal pattern recognition, and construction and optimization of deep learning models.

[0034] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0035] (1) Timestamp calibration is performed on multi-source raw data to obtain time-synchronized data, and spatial coordinate conversion is performed on the time-synchronized data to obtain unified projection data;

[0036] (2) Performing data quality assessment on the unified projection data to obtain quality labeled data;

[0037] (3) Perform multi-source data fusion on the quality-labeled data to obtain preliminary fused data;

[0038] (4) Eliminate noise from the preliminary fusion data using the Kalman filter algorithm to obtain noise-reduced optimized data;

[0039] (5) Adjust the temporal and spatial scale consistency of the noise reduction and optimization data to obtain data with unified scale;

[0040] (6) The uniform-scale data are discretized to obtain discretized time series data, and the discretized time series data are encoded with multidimensional features to obtain a standardized multidimensional time series precipitation dataset.

[0041] Specifically, the timestamp calibration is performed on the multi-source raw data to obtain time-synchronized data. The purpose of timestamp calibration is to ensure that data from different sources are compared and analyzed under the same time reference. The specific operation includes converting the timestamps of all data to a unified time format, usually using UTC (Coordinated Universal Time) as the standard time format. During the calibration process, it is necessary to deal with the time deviation caused by the clock error of different devices, and to perform time calibration through interpolation methods or timing alignment algorithms to ensure the consistency of the data in the time dimension. After obtaining the time-synchronized data, spatial coordinate conversion is performed to convert the data into a unified spatial reference system. The purpose of spatial coordinate conversion is to align data from different sources in space, and a geographic coordinate system (such as WGS84) is usually used as a unified projection reference system. During the conversion process, it is necessary to consider the transformation relationship between different coordinate systems, and to perform coordinate conversion through mathematical transformation formulas or GIS (Geographic Information System) tools to ensure the consistency of the data in the spatial dimension, and finally obtain unified projection data.

[0042] Based on the unified projection data, data quality assessment is performed to mark outliers or missing values ​​in the data and obtain quality-marked data. The purpose of data quality assessment is to ensure the reliability and accuracy of the data. Statistical analysis methods or machine learning algorithms are usually used to evaluate the data. Statistical analysis methods include calculating basic statistics of the data (such as mean, variance, extreme values, etc.) to identify obvious outliers; machine learning algorithms include anomaly detection algorithms (such as isolation forest algorithm, support vector machine, etc.) to identify abnormal patterns in the data through training models. After the quality-marked data is evaluated, outliers or missing values ​​are marked so that they can be corrected or eliminated in subsequent processing.

[0043] Multi-source data fusion is performed on the quality-marked data to obtain preliminary fused data. The purpose of multi-source data fusion is to comprehensively utilize data from different sources to improve the comprehensiveness and accuracy of the data. In the fusion process, the weighted average method or Bayesian fusion method is used to weight the multi-source data according to the weights to obtain the fused data. The weighted average method is to assign weights to data from different sources and weighted sum the data according to the weights; the Bayesian fusion method is based on the Bayesian probability model to fuse the data, consider the prior probability and likelihood function of the data, and obtain the probability distribution of the fused data. After the preliminary fused data is fused, noise is eliminated to improve the quality of the data. The preliminary fused data is denoised by the Kalman filter algorithm to obtain noise-reduced optimized data. The Kalman filter algorithm is a recursive estimation algorithm based on the state space model, which can effectively filter out the noise in the data and retain the effective information of the data. The specific operation includes constructing a state space model, setting the state transfer equation and the observation equation, and recursively filtering the data using the Kalman filter algorithm to obtain the optimized data. The advantage of the Kalman filter algorithm is that it can adaptively adjust the filter parameters, process non-stationary data, and has a better filtering effect.

[0044] After obtaining the noise reduction and optimization data, the spatiotemporal scale consistency adjustment is performed to obtain data with uniform scale. The purpose of spatiotemporal scale consistency adjustment is to standardize the data in time and space so that the data is consistent at different scales. Specific operations include time scale adjustment (such as sampling rate conversion, time window smoothing) and spatial scale adjustment (such as spatial interpolation, spatial resampling) of the data. Time scale adjustment makes the data consistent in time scale by changing the sampling frequency of the data or performing time window smoothing; spatial scale adjustment makes the data consistent in spatial scale by performing spatial interpolation or resampling of the data.

[0045] Data discretization is performed on the uniformly scaled data to obtain discretized time series data. The purpose of data discretization is to convert continuous data into discrete data for subsequent feature extraction and analysis. The specific operation includes equidistance discretization or equal-frequency discretization of the data to divide the data into several discrete intervals. Equidistance discretization is to divide the data according to fixed intervals to obtain discrete intervals; equal-frequency discretization is to divide the data according to frequency so that the amount of data contained in each interval is roughly the same.

[0046] Finally, the discretized time series data is encoded with multidimensional features to obtain a standardized multidimensional time series precipitation dataset. The purpose of multidimensional feature encoding is to convert discrete data into multidimensional feature vectors to facilitate subsequent spatiotemporal feature analysis. The specific operation includes feature extraction and encoding of the data, and converting the spatiotemporal features of the data into multidimensional feature vectors. Feature extraction methods include spatiotemporal sliding windows, symbolic dynamics encoding, phase space reconstruction, etc.; encoding methods include unique hot encoding, embedded vector encoding, etc. Through these methods, the spatiotemporal features of the data are encoded to form a multidimensional feature vector, and a standardized multidimensional time series precipitation dataset is obtained.

[0047] For example, when collecting multi-source original precipitation data in a certain area, including ground meteorological station observation data, satellite remote sensing data, and ground automatic weather station data. These data have different timestamps and spatial coordinates. Through timestamp calibration, all data are converted to a unified UTC time format. Then, through spatial coordinate conversion, the data is converted to a unified WGS84 coordinate system. The quality of the unified projection data is evaluated, and the outliers and missing values ​​in the data are identified using the isolation forest algorithm to obtain quality-marked data. Next, the quality-marked data is fused with multi-source data, and the weighted average method is used to weight the data from different sources according to the weight to obtain preliminary fused data. The preliminary fused data is denoised by the Kalman filter algorithm, and a state space model is constructed and recursively filtered to obtain denoised optimized data. The denoised optimized data is adjusted for temporal and spatial scale consistency. Through time window smoothing and spatial interpolation methods, the data is made consistent in time and space scales to obtain scale-unified data. Then, the scale-unified data is equidistantly discretized, and the data is divided into several discrete intervals to obtain discretized time series data. Finally, the discretized time series data is encoded with multidimensional features, and the spatiotemporal features of the data are converted into multidimensional feature vectors using symbolic dynamics encoding and phase space reconstruction methods to form a standardized multidimensional time series precipitation dataset. In this process, through steps such as timestamp calibration, spatial coordinate conversion, data quality assessment, multi-source data fusion, Kalman filtering, spatiotemporal scale consistency adjustment, data discretization and multidimensional feature encoding, the multi-source raw data is successfully processed into a standardized multidimensional time series precipitation dataset to ensure data quality and consistency.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] (1) Calculate the fractal dimension of the multidimensional time series precipitation dataset to obtain the spatiotemporal complexity characteristics;

[0050] (2) Perform symbolic dynamics encoding on the multidimensional time series precipitation dataset to obtain symbolic sequence features;

[0051] (3) Embed the multidimensional time-series precipitation dataset by reconstructing the phase space and obtain the phase space trajectory characteristics;

[0052] (4) Calculate the Lyapunov exponent of the phase space trajectory characteristics to obtain the system stability characteristics;

[0053] (5) Recursive quantitative analysis of multidimensional time series precipitation datasets was performed to obtain the periodic characteristics of the system;

[0054] (6) Perform multi-scale entropy analysis on the multidimensional time series precipitation dataset to obtain the time series complexity characteristics;

[0055] (7) Time-frequency analysis of multidimensional time series precipitation data sets was performed using Hilbert-Huang transform to obtain nonlinear time-frequency characteristics;

[0056] (8) Perform lagged mutual information analysis on multidimensional time series precipitation datasets to obtain spatiotemporal dependency characteristics;

[0057] (9) Perform nonlinear dimensionality reduction on the spatiotemporal complexity characteristics, symbol sequence characteristics, system stability characteristics, system periodicity characteristics, timing complexity characteristics, nonlinear time-frequency characteristics, and spatiotemporal dependency characteristics to obtain a set of precipitation spatiotemporal feature vectors.

[0058] Specifically, the fractal dimension of the multidimensional time series precipitation dataset is calculated to obtain the spatiotemporal complexity characteristics. Fractal dimension is a method of measuring complexity, which can reflect the complex structure of precipitation data in time and space. In the calculation process, the precipitation data is first segmented, and then the box counting method is used to calculate the fractal dimension of each segment. The box counting method embeds the data into boxes of different sizes and counts the number of data points in each box to calculate the fractal dimension. The larger the fractal dimension, the higher the spatiotemporal complexity of the data. The symbolic dynamics encoding of the multidimensional time series precipitation dataset is performed to obtain the symbolic sequence characteristics. Symbolic dynamics encoding is a method of converting a continuous data sequence into a discrete symbolic sequence, which can simplify the representation of the data and retain its main characteristics. In the encoding process, the data is first standardized, and then converted into a symbolic sequence according to a predefined threshold. For example, the threshold is set as three levels of high, medium and low precipitation, high precipitation is encoded as A, medium precipitation is encoded as B, and low precipitation is encoded as C, and finally a symbolic sequence consisting of A, B, and C is obtained.

[0059] By reconstructing the phase space and embedding the multidimensional time series precipitation data set, the phase space trajectory characteristics can be obtained. Phase space reconstruction is a nonlinear time series analysis method that can reveal the intrinsic dynamic characteristics of the data. In the reconstruction process, the delay time and embedding dimension are first determined, and then the original time series delay is embedded into the high-dimensional phase space to form a phase space trajectory. The delay time and embedding dimension can be determined by the autocorrelation function or the pseudo nearest neighbor algorithm. The phase space trajectory can reflect the dynamic behavior of the data and provide a basis for subsequent feature extraction. The Lyapunov exponent calculation of the phase space trajectory characteristics is to obtain the system stability characteristics. The Lyapunov exponent is an indicator of the stability of a dynamic system and can reflect the sensitivity of the system to the initial conditions. In the calculation process, several initial points are first selected on the phase space trajectory, and then the evolution trajectory of these points in the phase space is tracked to calculate the exponential separation speed between the trajectories. When the Lyapunov exponent is positive, the system is highly sensitive to the initial conditions and exhibits chaotic characteristics; when it is negative, the system tends to be stable.

[0060] The purpose of recursive quantitative analysis of multidimensional time series precipitation datasets is to obtain the periodic characteristics of the system. Recursive quantitative analysis is a tool for analyzing the periodicity of time series, which can reveal the periodic behavior in the data. During the analysis process, a recurrence plot is first constructed to map the time series into a two-dimensional matrix. The Euclidean distance between different time points is calculated to determine whether it is recursive. The periodic structure in the recurrence plot can reflect the periodic characteristics of the data, such as period length and period intensity.

[0061] The purpose of multi-scale entropy analysis on the multi-dimensional time series precipitation dataset is to obtain the time series complexity characteristics. Multi-scale entropy analysis is an indicator to measure the complexity of time series, which can capture the complexity of data at different scales. In the analysis process, the time series is first decomposed into sub-sequences of different time scales by multi-scale decomposition, such as wavelet transform or empirical mode decomposition, and then the entropy value of each sub-sequence is calculated. The larger the entropy value, the higher the complexity of the data and the greater the amount of information contained. The purpose of time-frequency analysis on the multi-dimensional time series precipitation dataset by Hilbert-Huang transform is to obtain nonlinear time-frequency characteristics. Hilbert-Huang transform is a nonlinear and non-stationary time series analysis method that can reveal the time-frequency characteristics of data. In the analysis process, the time series is first decomposed into several intrinsic mode functions (IMFs) by empirical mode decomposition, and then each IMF is Hilbert transformed to obtain the instantaneous frequency and instantaneous amplitude. By drawing the time-frequency diagram, the distribution of data at different times and frequencies can be intuitively observed.

[0062] The purpose of performing lagged mutual information analysis on multidimensional time series precipitation datasets is to obtain spatiotemporal dependency characteristics. Lagged mutual information analysis is a tool to measure the dependency between time series and can reveal the spatiotemporal association between data. During the analysis, the time series is first delayed and embedded to form a lagged series, and then the mutual information at different lag times is calculated to reflect the information transfer and dependency between the two time series. The larger the mutual information, the stronger the dependency between the two series.

[0063] Finally, nonlinear dimensionality reduction is performed on spatiotemporal complexity features, symbol sequence features, system stability features, system periodicity features, timing complexity features, nonlinear time-frequency features, and spatiotemporal dependency features to obtain a set of precipitation spatiotemporal feature vectors. The purpose of nonlinear dimensionality reduction is to reduce the dimension of features, retain the main information, and simplify the representation of data. Commonly used nonlinear dimensionality reduction methods include principal component analysis (PCA), t-distributed neighbor embedding (t-SNE), and autoencoders. In the dimensionality reduction process, the high-dimensional feature vector is first input into the dimensionality reduction algorithm, and the dimension of the feature is reduced by the optimization algorithm to obtain a low-dimensional feature representation. These low-dimensional feature vectors can retain most of the information of the original data, while reducing the computational complexity and improving the efficiency of subsequent analysis.

[0064] For example, a multidimensional time series precipitation data set of a certain area is collected, and the fractal dimension of the data is calculated. The fractal dimension of each time period is calculated using the box counting method to obtain the spatiotemporal complexity characteristics. Then, the data is symbolically coded, and three levels of precipitation, high, medium, and low, are set to convert the data into a symbolic sequence. Then, by reconstructing the phase space, the time series delay is embedded into the three-dimensional phase space to form a phase space trajectory. The Lyapunov exponent is calculated for the phase space trajectory, and the initial point is selected to track its trajectory, and the exponential separation speed is calculated to obtain the system stability characteristics. Then, the data is recursively quantitatively analyzed, a recursive graph is constructed, and the recursive relationship between time points is determined by calculating the Euclidean distance to obtain the system periodicity characteristics. Then, the data is multi-scale entropy analysis is performed, and the data is decomposed into subsequences of different time scales through wavelet transform. The entropy value of each subsequence is calculated to obtain the time series complexity characteristics. Through the Hilbert-Huang transform, the data is subjected to empirical mode decomposition to obtain several intrinsic mode functions. Each IMF is subjected to a Hilbert transform to obtain the instantaneous frequency and instantaneous amplitude, forming a time-frequency diagram and obtaining nonlinear time-frequency features. Finally, the data is subjected to lagged mutual information analysis, and a lagged sequence is formed by delayed embedding. The mutual information at different lag times is calculated to obtain the spatiotemporal dependency features. These features are subjected to nonlinear dimensionality reduction, and the principal component analysis method is used to reduce the high-dimensional feature vectors to three-dimensional feature vectors to form a precipitation spatiotemporal feature vector set. In this process, a series of feature extraction and analysis methods are used to successfully convert the multidimensional time series precipitation data set into a precipitation spatiotemporal feature vector set.

[0065] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0066] (1) Performing spatiotemporal sliding window segmentation on the precipitation spatiotemporal feature vector set to obtain a multi-scale feature segment set;

[0067] (2) Perform spatiotemporal pattern coding on the multi-scale feature fragment set to obtain a pattern coding sequence;

[0068] (3) Perform frequent pattern mining on the pattern coding sequence to obtain a set of candidate spatiotemporal patterns;

[0069] (4) Analyze the spatiotemporal association rules of the candidate spatiotemporal pattern set to obtain the spatiotemporal association pattern;

[0070] (5) Cluster the spatiotemporal correlation patterns using the spectral clustering algorithm to obtain preliminary clustering results;

[0071] (6) Calculate the fuzzy membership degree of the preliminary clustering results to obtain soft clustering assignments;

[0072] (7) Perform density peak detection on the soft clustering assignment to obtain the cluster center;

[0073] (8) Redistribute the precipitation spatiotemporal feature vector set based on the cluster center to obtain the clustering result;

[0074] (9) Perform outlier detection on the clustering results to obtain a candidate set of abnormal precipitation patterns;

[0075] (10) quantifying the abnormal precipitation pattern candidate set through the isolation forest algorithm to obtain the abnormal precipitation pattern list;

[0076] (11) Integrate the spatiotemporal correlation patterns, clustering results, and abnormal precipitation pattern lists to obtain a precipitation spatiotemporal pattern library.

[0077] Specifically, performing spatiotemporal sliding window segmentation on the precipitation spatiotemporal feature vector set is a key step to obtain a multi-scale feature fragment set. Spatiotemporal sliding window segmentation is to divide the original spatiotemporal feature vector set into fixed-size fragments to facilitate the capture of multi-scale spatiotemporal features. Specifically, a fixed time window length and space window size are set, and then the window is slid with a certain step length in the entire spatiotemporal data set to extract the feature data in the window. For example, the time window length is set to 10 days, the space window is 10 square kilometers, and the step length is 1 day and 1 square kilometer. Each slide extracts the data in the window to form a multi-scale feature fragment set. Spatiotemporal pattern coding of the multi-scale feature fragment set is to convert complex spatiotemporal features into discrete patterns that are easy to analyze. Spatiotemporal pattern coding encodes each fragment into a pattern label by defining several feature patterns. For example, the precipitation intensity can be divided into several levels (such as high, medium, and low), and then encoded according to the precipitation intensity of each fragment to obtain a pattern coding sequence. This encoding process can simplify continuous spatiotemporal feature data into discrete pattern labels, facilitating subsequent pattern mining and analysis.

[0078] Frequent pattern mining of pattern coding sequences is to discover spatiotemporal patterns that frequently appear in data. Frequent pattern mining uses the Apriori algorithm or the FP-growth algorithm to mine a set of patterns that frequently appear in a data set. The specific operation is to scan the coding sequence, count the frequency of each pattern and its combination, set a support threshold, filter out patterns whose frequency exceeds the threshold, and form a candidate spatiotemporal pattern set. For example, during the analysis process, it is found that a certain precipitation pattern (such as high precipitation for 3 consecutive days) frequently appears in the data set. Such a pattern is a candidate spatiotemporal pattern.

[0079] The purpose of performing spatiotemporal association rule analysis on the candidate spatiotemporal pattern set is to reveal the association relationship between patterns. Spatiotemporal association rule analysis identifies pattern pairs with significant associations by calculating the support, confidence, and lift between patterns. The specific operation is to perform statistics on each candidate pattern pair, calculate the frequency of their simultaneous occurrence (support), the probability of the latter pattern appearing when the previous pattern appears (confidence), and the strength of association (lift). For example, through analysis, it is found that there is a high confidence and high lift association relationship between a high precipitation pattern and a low precipitation pattern, which can be identified as a spatiotemporal association pattern.

[0080] The purpose of clustering spatiotemporal correlation patterns by using the spectral clustering algorithm is to group similar patterns together. The spectral clustering algorithm is a clustering method based on graph theory. It constructs a similarity matrix between correlation patterns, performs eigenvalue decomposition on the matrix, projects the pattern into a low-dimensional space, and then performs K-means clustering in the low-dimensional space. The specific operation is to first calculate the similarity between each pair of spatiotemporal correlation patterns, construct a similarity matrix, then perform eigenvalue decomposition on the matrix, select the first k eigenvectors as the new feature space, perform K-means clustering on the feature space, and obtain preliminary clustering results. For example, through the spectral clustering algorithm, several high precipitation patterns with similar spatiotemporal characteristics are clustered in one cluster to form a preliminary clustering result.

[0081] The purpose of performing fuzzy membership calculation on the preliminary clustering results is to further refine the clustering results. The fuzzy membership calculation uses the fuzzy C-means algorithm to calculate the probability (membership) of each data point belonging to each cluster. The specific operation is to first initialize the center of each cluster, then calculate the membership based on the distance between the data point and the cluster center, update the position of the cluster center, and iterate until convergence. In this way, the soft clustering assignment of each data point can be obtained. For example, in the preliminary clustering results, a data point may belong to multiple clusters at the same time. Through the fuzzy membership calculation, the probability distribution of it belonging to each cluster can be obtained.

[0082] The purpose of performing density peak detection on soft cluster assignment is to find the cluster center. Density peak detection identifies points with high density and far distance as cluster centers by calculating the local density of each data point and the distance to the nearest high-density point. The specific operation is to first calculate the local density of each point, usually using a Gaussian kernel function, then calculate the distance from each point to the nearest high-density point, draw a density-distance graph, and identify the cluster center by observing the density peak. For example, through density peak detection, several points with high density and far distance from each other are identified in the soft cluster assignment as the final cluster centers.

[0083] The purpose of redistributing the precipitation spatiotemporal feature vector set based on the cluster center is to cluster the results. The redistribution process is to assign each data point to the cluster where the nearest cluster center is located to form a new clustering result. The specific operation is to calculate the distance from each data point to all cluster centers and assign the data point to the cluster to which the nearest cluster center belongs. For example, during the redistribution process, a data point that originally belonged to multiple clusters may be assigned to the cluster where the nearest cluster center is located, thereby forming an optimized clustering result.

[0084] Outlier detection is performed on the clustering results to identify abnormal precipitation patterns. Outlier detection calculates the distance between each data point and the center of its cluster and identifies points that are farther away as outliers. The specific operation is to first calculate the distance from each data point to the cluster center, and then set a threshold. Points exceeding the threshold are outliers, forming a candidate set of abnormal precipitation patterns. For example, in the clustering results, a data point that is far from the cluster center is identified as an outlier and a candidate point for abnormal precipitation patterns.

[0085] The purpose of quantifying the abnormal precipitation pattern candidate set through the isolation forest algorithm is to finally determine the abnormal precipitation pattern list. The isolation forest algorithm is an anomaly detection algorithm based on decision trees. It randomly divides the data points by constructing multiple random trees, and calculates the isolation path length of each point in the tree. The shorter the isolation path, the more likely it is an abnormal point. The specific operation is to construct multiple random trees, calculate the isolation path length of each abnormal precipitation pattern candidate point in the tree, set a threshold, and identify the points with isolation path length shorter than the threshold as abnormal points to form an abnormal precipitation pattern list. For example, through the isolation forest algorithm, the points in the abnormal precipitation pattern candidate set are calculated, and finally several abnormal precipitation patterns are identified to form an abnormal precipitation pattern list.

[0086] The purpose of integrating spatiotemporal correlation patterns, clustering results, and abnormal precipitation pattern lists is to build a complete precipitation spatiotemporal pattern library. The spatiotemporal correlation patterns reveal the correlation relationships in precipitation data, the clustering results provide clustering information of precipitation patterns, and the abnormal precipitation pattern list identifies abnormal patterns in precipitation data. By integrating this information, a comprehensive precipitation spatiotemporal pattern library is formed, which provides a basis for subsequent precipitation prediction and analysis. For example, in the integration process, the discovered spatiotemporal correlation patterns, the optimized clustering results, and the identified abnormal precipitation patterns are combined to form a precipitation spatiotemporal pattern library containing a variety of information, providing comprehensive support for further analysis and application.

[0087] For example, a multidimensional time series precipitation data set for a certain area is collected. First, the data is segmented into a time-space sliding window. The time window length is set to 10 days and the space window is set to 10 square kilometers. The data is slid with a step size of 1 day and 1 square kilometer. The data in the window is extracted to form a multi-scale feature fragment set. Then, the multi-scale feature fragment set is coded with a time-space pattern. The precipitation intensity is divided into three levels: high, medium, and low. The fragments are coded according to the precipitation intensity to obtain a pattern coding sequence. Next, the pattern coding sequence is mined for frequent patterns. The Apriori algorithm is used to mine the frequently occurring precipitation patterns to form a candidate time-space pattern set. The candidate time-space pattern set is analyzed for time-space association rules, the support, confidence, and lift between the patterns are calculated, and the pattern pairs with significant associations are identified to obtain time-space association patterns. The time-space association patterns are clustered by the spectral clustering algorithm, the similarity matrix is ​​constructed, and the eigenvalue decomposition and K-means clustering are performed to obtain preliminary clustering results. The preliminary clustering results are fuzzy membership calculations, and the fuzzy C-means algorithm is used to calculate the soft clustering assignment of each data point. Density peak detection is performed on the soft clustering assignment, local density and distance are calculated, and cluster centers are identified. The precipitation spatiotemporal feature vector set is redistributed based on the cluster centers, and the distance from each data point to the cluster center is calculated to form a clustering result. Outlier detection is performed on the clustering results, the distance from each data point to the cluster center is calculated, outliers are identified, and a candidate set of abnormal precipitation patterns is formed. The abnormal precipitation pattern candidate set is quantified by the isolation forest algorithm, the isolation path length is calculated, the abnormal points are identified, and a list of abnormal precipitation patterns is formed. Finally, the spatiotemporal correlation patterns, clustering results, and abnormal precipitation pattern lists are integrated to build a complete precipitation spatiotemporal pattern library.

[0088] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0089] (1) Embed the precipitation spatiotemporal pattern library to obtain the pattern embedding matrix;

[0090] (2) Perform feature mapping on the clustering results to obtain the clustering feature tensor;

[0091] (3) Encode the abnormal pattern of the abnormal precipitation pattern list to obtain the abnormal feature vector;

[0092] (4) Perform multimodal feature fusion on the pattern embedding matrix, clustering feature tensor and abnormal feature vector to obtain fused feature representation data;

[0093] (5) Use graph convolutional networks to model the spatiotemporal dependency of the fused feature representation data and obtain a spatiotemporal relationship graph;

[0094] (6) Calculate the attention mechanism on the spatiotemporal relationship graph to obtain weighted spatiotemporal features;

[0095] (7) Use a bidirectional long short-term memory network to perform sequence modeling on the weighted spatiotemporal features and obtain time series prediction results;

[0096] (8) Perform deconvolution operation on the time series prediction results to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model.

[0097] Specifically, the precipitation spatiotemporal pattern library is pattern embedded to obtain a pattern embedding matrix. Pattern embedding is a method of converting discrete spatiotemporal patterns into continuous vector representations. The specific operation is to represent each spatiotemporal pattern as a high-dimensional vector, which is usually trained by word embedding algorithms (such as Word2Vec or GloVe). These algorithms build a large spatiotemporal pattern dictionary and train vector representations based on the co-occurrence relationship of spatiotemporal patterns. For example, in Word2Vec, each spatiotemporal pattern is embedded in a continuous vector space so that similar spatiotemporal patterns are closer in the vector space, thereby forming a pattern embedding matrix. The clustering results are feature mapped to obtain a clustering feature tensor. Feature mapping is the process of representing each cluster in the clustering results as a feature vector. The specific operation is to first determine the center point of each cluster, then calculate the distance from each data point to the center of the cluster to which it belongs, and use these distances as components of the feature vector. In this way, the clustering results can be converted into a high-dimensional feature tensor, in which each element in the tensor represents the distance from the data point to the center of the cluster. This feature representation method can retain important information in the clustering results, facilitating subsequent analysis and processing.

[0098] Abnormal pattern encoding is performed on the abnormal precipitation pattern list to obtain an abnormal feature vector. Abnormal pattern encoding is the process of representing each abnormal pattern in the abnormal precipitation pattern list as a feature vector. The specific operation is to first standardize each abnormal pattern and then use the standardized eigenvalue as the component of the feature vector. In this way, the abnormal precipitation pattern list can be converted into a high-dimensional feature vector, each of which represents an abnormal pattern. This encoding method can retain important information in the abnormal pattern, which is convenient for subsequent analysis and processing. Multimodal feature fusion is performed on the pattern embedding matrix, clustering feature tensor and abnormal feature vector to obtain fused feature representation data. Multimodal feature fusion is a method of combining feature representations from different sources to form a comprehensive feature representation. The specific operation is to splice the pattern embedding matrix, clustering feature tensor and abnormal feature vector, and then fuse the spliced ​​features through a fully connected neural network to obtain a new feature representation. This process can retain important information in different modal features while eliminating redundant information to form a comprehensive feature representation data, which is convenient for subsequent spatiotemporal dependency modeling.

[0099] The fused feature representation data is modeled for spatiotemporal dependency through a graph convolutional network (GCN) to obtain a spatiotemporal relationship graph. A graph convolutional network is a neural network based on a graph structure that can effectively capture the spatiotemporal dependency in the data. The specific operation is to first construct the fused feature representation data into a graph structure, in which the nodes represent the spatiotemporal features and the edges represent the spatiotemporal dependency between the nodes. Then, the graph structure is convolved by GCN to extract the features of each node. The convolution operation includes calculating the weighted sum of the features of each node and its neighboring nodes, and then obtaining a new feature representation of the node through a nonlinear activation function. By superimposing multiple layers of GCN, the deep features in the spatiotemporal relationship graph can be extracted layer by layer to form a comprehensive spatiotemporal relationship graph. The attention mechanism is calculated on the spatiotemporal relationship graph to obtain weighted spatiotemporal features. The attention mechanism is a method of assigning importance weights to different features, which can improve the effect of feature representation. The specific operation is to first calculate the attention weights of the neighboring nodes for each node in the spatiotemporal relationship graph. These weights can be calculated by an attention function (such as dot product attention or multi-head attention). Then, the features of each node are weighted averaged, and the weights are calculated by the attention function to form weighted spatiotemporal features. This method can highlight the important features in the spatiotemporal relationship graph and improve the effect of feature representation.

[0100] The weighted spatiotemporal features are sequence modeled through a bidirectional long short-term memory network (BiLSTM) to obtain the time series prediction results. The bidirectional long short-term memory network is a neural network that can process time series data and can capture the forward and backward dependencies in the sequence data. The specific operation is to input the weighted spatiotemporal features into the BiLSTM network. The BiLSTM network consists of two LSTM layers, the forward LSTM layer processes the forward dependencies of the sequence data, and the backward LSTM layer processes the reverse dependencies of the sequence data. Through the combination of the forward and backward LSTM layers, the global dependencies in the sequence data can be captured and the time series prediction results can be output. The time series prediction results are deconvolved to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model. The deconvolution operation is a method of restoring the spatial structure of data, which can map low-dimensional features back to the original high-dimensional space. The specific operation is to input the time series prediction results into the deconvolution layer. The deconvolution layer maps the low-dimensional features back to the original high-dimensional space through the reverse convolution operation, thereby forming an end-to-end precipitation prediction model. This method can effectively restore the spatial structure of the original data and improve the effect of the prediction model.

[0101] For example, to collect the spatiotemporal precipitation data of a certain area, firstly embed the data into patterns and obtain the pattern embedding matrix through Word2Vec algorithm training. Then, feature mapping is performed on the clustering results, and the distance from each data point to the cluster center is calculated to form a cluster feature tensor. Then, the abnormal pattern encoding is performed on the abnormal precipitation pattern list, and each abnormal pattern is converted into a feature vector. The pattern embedding matrix, clustering feature tensor and abnormal feature vector are spliced, and multimodal feature fusion is performed through a fully connected neural network to obtain fused feature representation data. The spatiotemporal dependency of the fused feature representation data is modeled through a graph convolutional network, a spatiotemporal relationship graph is constructed, and deep features in the spatiotemporal relationship graph are extracted through a multi-layer GCN. The attention mechanism is calculated for the spatiotemporal relationship graph, and feature importance weights are assigned to form weighted spatiotemporal features. The weighted spatiotemporal features are input into a bidirectional long short-term memory network, and the global dependency of the sequence data is captured through the forward and backward LSTM layers to output the time series prediction results. Finally, a deconvolution operation is performed on the time series prediction results to map the low-dimensional features back to the high-dimensional space to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model.

[0102] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0103] (1) Segment the data stream of real-time precipitation observation data to obtain time series data segments;

[0104] (2) Conduct concept drift detection on time series data segments to obtain data distribution change indicators;

[0105] (3) Based on the data distribution change index, the parameter sensitivity analysis of the global atmospheric reanalysis precipitation fusion prediction model was carried out to obtain the sensitive parameter set;

[0106] (4) Incrementally learn the sensitive parameter set to obtain an updated parameter vector;

[0107] (5) Fusing the updated parameter vector with the global atmospheric reanalysis precipitation fusion prediction model to obtain a temporary prediction model;

[0108] (6) compressing the temporary prediction model to obtain a lightweight prediction model;

[0109] (7) Distributed training of lightweight prediction models is performed through federated learning to obtain a multi-source collaborative optimization model;

[0110] (8) Perform model distillation on the multi-source collaborative optimization model to obtain a knowledge transfer model;

[0111] (9) Conduct adversarial sample training on the knowledge transfer model to obtain a robustness enhanced model;

[0112] (10) Multi-task learning is performed on the robustness enhancement model to obtain the target precipitation spatiotemporal prediction model.

[0113] Specifically, data stream segmentation of real-time precipitation observation data to obtain time series data segments is the first step to achieve dynamic update and online learning. Data stream segmentation is to segment continuous real-time precipitation observation data according to fixed time windows to form several time series data segments. These segments can be time windows of fixed length, such as hourly, daily or weekly data segments. By processing these data segments, the dynamic changes of precipitation data can be captured. For example, the real-time precipitation observation data is segmented into a time series data segment every hour, so that a continuous set of time series data segments can be obtained. Concept drift detection of time series data segments to obtain data distribution change indicators is an important step in detecting changes in precipitation data distribution. Concept drift detection is a method for identifying changes in data distribution, which can detect trend changes, seasonal changes or sudden changes in data. Commonly used concept drift detection algorithms include Page-Hinkley detection method, ADWIN algorithm and DDM algorithm. The specific operation is to perform statistical analysis on each time series data segment, calculate the mean, variance and other statistics of the data, and then compare them with historical data to identify changes in data distribution. For example, through the Page-Hinkley detection method, the mean and variance of each time series data segment can be calculated and compared with historical data to identify the changing trend of data distribution and obtain the data distribution change index.

[0114] Based on the data distribution change index, the parameter sensitivity analysis of the global atmospheric reanalysis precipitation fusion prediction model is carried out. Obtaining the sensitive parameter set is a key step in optimizing the prediction model. Parameter sensitivity analysis is a method to evaluate the impact of model parameters on prediction results. By analyzing the impact of different parameters on the model output, the sensitive parameters that have a greater impact on model performance are identified. The specific operation is to perturb each parameter in the model, calculate the change amplitude of the model output before and after the perturbation, and then evaluate the sensitivity of the parameters based on the change amplitude. For example, by making a small perturbation to each parameter in the precipitation prediction model and calculating the change amplitude of the model output, the sensitive parameters that have a greater impact on the model performance can be identified to form a sensitive parameter set.

[0115] Incremental learning of the sensitive parameter set to obtain the updated parameter vector is an important step in model adaptive learning. Incremental learning is a method of gradually updating model parameters by continuously learning new data and gradually optimizing model parameters. The specific operation is to take the parameters in the sensitive parameter set as the learning object, make a small adjustment to each parameter, calculate the error between the adjusted model output and the actual observed data, adjust the parameter value according to the error, and iterate until the error is minimized. For example, by making a small adjustment to the sensitive parameters, calculating the error between the model output and the actual observed data, adjusting the parameter value according to the error, and forming an updated parameter vector.

[0116] The key step to achieve dynamic model update is to fuse the updated parameter vector with the global atmospheric reanalysis precipitation fusion prediction model to obtain a temporary prediction model. Model fusion is a method of combining new and old model parameters to form a new model by fusing the updated parameter vector with the original model parameters. The specific operation is to perform a weighted average of the updated parameter vector and the original model parameters according to certain weights to obtain a new parameter vector, and then apply the new parameter vector to the model to form a temporary prediction model. For example, by performing a weighted average of the updated parameter vector and the original model parameters according to a weight of 0.5, a new parameter vector is obtained and applied to the precipitation prediction model to form a temporary prediction model.

[0117] Compressing the temporary prediction model to obtain a lightweight prediction model is a key step to improve the computational efficiency of the model. Model compression is a method to reduce the complexity of the model. It reduces the computational complexity and storage requirements of the model through techniques such as pruning, quantization, or distillation. The specific operation is to prune the temporary prediction model to delete unimportant parameters and connections in the model, or quantize the model parameters to convert the parameters represented by floating point numbers into low-precision representations, or transfer the knowledge of the complex model to a simple model through model distillation. For example, by pruning the temporary prediction model, deleting unimportant parameters and connections, reducing the computational complexity of the model, and forming a lightweight prediction model.

[0118] Distributed training of lightweight prediction models through federated learning to obtain a multi-source collaborative optimization model is a key step in achieving collaborative optimization of multi-source data. Federated learning is a distributed machine learning method that achieves collaborative optimization of data by training models on multiple nodes. The specific operation is to distribute the lightweight prediction model to multiple nodes, each node uses local data to train the model, and then aggregate the training results to the central server to perform weighted averaging of the model parameters to form a globally optimized model. For example, by distributing the lightweight prediction model to multiple meteorological observation stations, each observation station uses local data for training, and then aggregates the training results to the central server to perform weighted averaging of the model parameters to form a multi-source collaborative optimization model.

[0119] Performing model distillation on the multi-source collaborative optimization model to obtain a knowledge transfer model is a key step to improve the model's lightweight and performance. Model distillation is a method of transferring the knowledge of a complex model to a simple model, and improving the performance of the simple model by allowing the simple model to learn the output of the complex model. The specific operation is to first use the multi-source collaborative optimization model to predict the training data to obtain the prediction results, and then let the knowledge transfer model learn these prediction results, and optimize the parameters of the knowledge transfer model by minimizing the error between the prediction results and the actual observed data. For example, by using the multi-source collaborative optimization model to predict the training data to obtain the prediction results, and then let the knowledge transfer model learn these prediction results, and form a knowledge transfer model by minimizing the error between the prediction results and the actual observed data.

[0120] Training the knowledge transfer model with adversarial samples to obtain a robustness-enhanced model is a key step in improving the model's anti-interference ability. Adversarial sample training is a method to improve the robustness of the model. By adding adversarial samples during the training process, the model's resistance to noise and attacks is improved. The specific operation is to first generate adversarial samples, which are generated by making slight perturbations to the original data, and then add the adversarial samples to the training data for model training. By minimizing the prediction error between the adversarial samples and the original data, the robustness of the model is improved. For example, by making slight perturbations to the original precipitation data, adversarial samples are generated, and then the adversarial samples are added to the training data for model training to form a robustness-enhanced model.

[0121] Performing multi-task learning on the robustness enhancement model to obtain the target precipitation spatiotemporal prediction model is a key step to improve the overall performance of the model. Multi-task learning is a method of optimizing multiple tasks simultaneously, and by sharing model parameters, the collaborative optimization of multiple tasks is achieved. The specific operation is to add tasks such as precipitation prediction, precipitation anomaly detection, and precipitation pattern recognition to the model training at the same time, and by sharing model parameters, jointly optimize to improve the overall performance of the model. For example, by adding tasks such as precipitation prediction, precipitation anomaly detection, and precipitation pattern recognition to the model training at the same time, and by sharing model parameters, jointly optimize, a target precipitation spatiotemporal prediction model is formed.

[0122] For example, firstly, the real-time precipitation observation data of a certain area is segmented into a time series data segment every hour. Then, the concept drift detection is performed on these time series data segments, and the mean and variance of each segment are calculated by using the Page-Hinkley detection method. The data distribution changes are compared with the historical data to identify the changes in the data distribution and obtain the data distribution change indicators. Based on these change indicators, the parameter sensitivity analysis of the global atmospheric reanalysis precipitation fusion prediction model is performed, and the model parameters are slightly perturbed to calculate the model output changes and identify the sensitive parameter set. The sensitive parameter set is incrementally learned, and the error between the model output and the observed data is minimized by adjusting the parameter value to obtain the updated parameter vector. The updated parameter vector is weighted averaged with the original model parameters to form a temporary prediction model. The temporary prediction model is pruned to delete unimportant parameters to form a lightweight prediction model. The lightweight model is distributed to multiple observation stations through federated learning, trained with local data, and the training results are aggregated to form a multi-source collaborative optimization model. The output of the multi-source collaborative optimization model is used to train the knowledge transfer model and form a knowledge transfer model by minimizing the error. The knowledge transfer model is trained with adversarial samples to generate adversarial samples, improve the model robustness, and form a robustness enhanced model. Finally, through multi-task learning, precipitation prediction, anomaly detection, and pattern recognition tasks are added to the training at the same time to form a target precipitation spatiotemporal prediction model.

[0123] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0124] (1) Perform spatiotemporal decomposition on the output of the target precipitation spatiotemporal prediction model to obtain multi-scale precipitation prediction results;

[0125] (2) Perform spatiotemporal autocorrelation analysis on multi-scale precipitation forecast results to obtain the evolution characteristics of precipitation patterns;

[0126] (3) Perform topological data analysis on the evolution characteristics of precipitation patterns to obtain the topological structure of precipitation patterns;

[0127] (4) Identify extreme value events based on the precipitation pattern topology and obtain extreme precipitation risk assessment;

[0128] (5) Conduct scenario simulation for extreme precipitation risk assessment and obtain multi-scenario precipitation impact predictions;

[0129] (6) Conduct system vulnerability assessment on multi-scenario precipitation impact prediction through complex network analysis and obtain key influencing factors;

[0130] (7) Carry out causal inference on key influencing factors and obtain the dynamic model of precipitation system;

[0131] (8) Based on the precipitation system dynamics model, a spatiotemporal knowledge graph is constructed to obtain a precipitation knowledge network;

[0132] (9) Conduct interpretability analysis on the precipitation knowledge network and obtain a global atmospheric reanalysis precipitation fusion data analysis report;

[0133] (10) Based on the global atmospheric reanalysis precipitation fusion data analysis report, resource optimization allocation is carried out to obtain the weather processing pipeline scheduling strategy.

[0134] Specifically, the spatiotemporal decomposition of the output of the target precipitation spatiotemporal prediction model is achieved through multi-scale wavelet transform. This process decomposes the precipitation prediction results into components of different spatiotemporal scales, including long-term trends, seasonal changes, and short-term fluctuations. For example, for the annual precipitation forecast of a region, a 10-year growth trend, 3-5-year periodic fluctuations, and seasonal changes within the year may be obtained. The spatiotemporal autocorrelation analysis of these multi-scale precipitation prediction results is performed using the Moran's I index and time series autocorrelation function. The Moran's I index is used to quantify the autocorrelation in space, while the time series autocorrelation function is used to identify temporal dependencies. This step can reveal the temporal and spatial evolution characteristics of precipitation patterns, such as the movement trend of precipitation belts or the periodic changes in local precipitation intensity.

[0135] For the identified precipitation pattern evolution characteristics, topological data analysis is performed. This step uses the persistence homology algorithm to represent the precipitation pattern as a high-dimensional topological structure. By calculating the Betti number and persistence graph, the topological characteristics of the precipitation pattern, such as connectivity and holes, can be captured. These topological characteristics reflect the structural stability and complexity of the precipitation system. Based on the obtained precipitation pattern topological structure, extreme value theory is used to identify extreme value events. This process uses the generalized extreme value distribution (GEV) and peak overthreshold (POT) model to model precipitation events that exceed a specific threshold. In this way, the probability and intensity of extreme precipitation events can be evaluated, thereby obtaining extreme precipitation risk assessment results.

[0136] The extreme precipitation risk assessment results were simulated and a variety of possible precipitation scenarios were generated using the Monte Carlo method. These scenarios took into account different climate change pathways, land use changes and human activity impacts, thereby obtaining multi-scenario precipitation impact forecasts.

[0137] The system vulnerability assessment of multi-scenario precipitation impact forecasts was performed through complex network analysis. This step constructed a precipitation impact network, in which nodes represent different geographical areas or system components, and edges represent the mutual influence between them. By calculating indicators such as network centrality and clustering coefficient, the key influencing factors in the system were identified. Causal inference was performed on these key influencing factors using structural equation models (SEM) or Bayesian networks. This process established a causal relationship diagram between factors, resulting in a precipitation system dynamics model. The model describes how different factors interact and affect the dynamic changes of the precipitation system.

[0138] Based on the precipitation system dynamics model, a spatiotemporal knowledge graph is constructed. This step maps the entities and relationships in the dynamics model into the graph database to form a knowledge network containing temporal and spatial information. The network contains not only precipitation-related entities (such as climate factors and geographical features), but also the spatiotemporal relationships between them.

[0139] The constructed precipitation knowledge network is analyzed for interpretability using techniques such as SHAP (SHapley AdditiveexPlanations) values. This process can explain the prediction results of the model and reveal the contribution of different factors to precipitation prediction, thereby generating a comprehensive global atmospheric reanalysis precipitation fusion data analysis report. Finally, based on this data analysis report, resource optimization is performed. This step uses multi-objective optimization algorithms, such as genetic algorithms or particle swarm optimization, to balance the meteorological observation needs, forecast accuracy requirements, and resource constraints in different regions, and finally obtains the weather processing pipeline scheduling strategy.

[0140] For example, in a large basin spanning multiple climate zones, 50 years of historical precipitation data and forecasts for the next 30 years were analyzed using the above methods. Spatiotemporal decomposition showed the long-term trend of annual precipitation in different regions, such as an increasing trend in the upstream area and a decreasing trend in the downstream area. Spatiotemporal autocorrelation analysis revealed the seasonal migration characteristics of precipitation patterns, and topological data analysis identified several key precipitation "hotspots". Extreme event analysis predicted that in the next 30 years, some areas in the basin may face a 30% increase in the risk of once-in-a-century rainstorms. Scenario simulation and complex network analysis further pointed out that this increased risk is mainly affected by regional land use changes and global warming. Causal inference models quantified the relative importance of these factors and provided a basis for the development of targeted adaptation strategies. Finally, based on these analysis results, the layout of meteorological observation stations and the resource allocation of early warning systems were optimized, and the forecasting and response capabilities for extreme precipitation events were improved.

[0141] The above describes the time series feature analysis method based on multidimensional data in the embodiment of the present application. The following describes the time series feature analysis system based on multidimensional data in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the time series feature analysis system based on multidimensional data includes:

[0142] The acquisition module 201 is used to collect and preprocess the multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations to obtain a standardized multi-dimensional time series precipitation data set;

[0143] A dimension reduction module 202 is used to extract spatiotemporal features and perform dimension reduction processing on the multi-dimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set;

[0144] The clustering module 203 is used to perform multi-scale spatiotemporal pattern recognition and cluster analysis on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results and an abnormal precipitation pattern list;

[0145] A construction module 204 is used to construct a deep learning precipitation prediction model based on the precipitation spatiotemporal pattern library, the clustering results and the abnormal precipitation pattern list to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model;

[0146] An updating module 205 is used to dynamically update and online learn the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model;

[0147] The analysis module 206 is used to perform multi-dimensional spatiotemporal analysis on the target precipitation spatiotemporal prediction model to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy.

[0148] Through the collaborative cooperation of the above components, a standardized multi-dimensional time series precipitation dataset can be obtained by collecting and preprocessing multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations. This dataset combines the advantages of multi-source data, avoids the limitations of a single data source, and makes the data more comprehensive and accurate. On this basis, the multi-dimensional time series precipitation dataset is subjected to spatiotemporal feature extraction and dimensionality reduction processing to obtain a precipitation spatiotemporal feature vector set. In this way, the complex spatiotemporal characteristics of precipitation can be effectively captured, and the data expression ability and processing efficiency can be improved. Secondly, multi-scale spatiotemporal pattern recognition and clustering analysis are performed on the precipitation spatiotemporal feature vector set, and a precipitation spatiotemporal pattern library, clustering results and a list of abnormal precipitation patterns can be constructed. This not only helps to deeply explore the spatiotemporal pattern of precipitation, but also can identify abnormal precipitation patterns and improve the prediction ability of extreme weather events. In addition, based on the precipitation spatiotemporal pattern library, clustering results and abnormal precipitation pattern list, a deep learning precipitation prediction model is constructed to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model. This deep learning model can effectively use large-scale data for training, capture the complex nonlinear characteristics of precipitation, and significantly improve the accuracy of precipitation forecasts. Through dynamic updating and online learning, the global atmospheric reanalysis precipitation fusion prediction model is optimized to obtain the target precipitation spatiotemporal prediction model, so that the model can adaptively adjust parameters and respond to weather changes in a timely manner, thereby improving the real-time and accuracy of the forecast. Furthermore, a multi-dimensional spatiotemporal analysis is performed on the target precipitation spatiotemporal prediction model, and a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy can be generated. The precipitation forecast results can be comprehensively evaluated and analyzed. The present invention not only significantly improves the accuracy and real-time nature of precipitation forecasts, but also enhances the prediction ability of extreme weather events through the fusion of multi-source data, spatiotemporal feature extraction, multi-scale spatiotemporal pattern recognition, and construction and optimization of deep learning models.

[0149] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0150] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A time series feature analysis method based on multidimensional data, characterized in that: The time series feature analysis method based on multidimensional data includes: Collect and preprocess the multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations to obtain a standardized multi-dimensional time series precipitation dataset; Performing spatiotemporal feature extraction and dimensionality reduction processing on the multidimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set; Performing multi-scale spatiotemporal pattern recognition and cluster analysis on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results, and an abnormal precipitation pattern list; Based on the precipitation spatiotemporal pattern library, the clustering results and the abnormal precipitation pattern list, a deep learning precipitation prediction model is constructed to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model; Dynamically updating and online learning the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model; A multi-dimensional spatiotemporal analysis is performed on the target precipitation spatiotemporal prediction model to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy.

2. The method for analyzing time series characteristics based on multidimensional data according to claim 1, characterized in that: The multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations are collected and preprocessed to obtain a standardized multi-dimensional time series precipitation dataset, including: Performing time stamp calibration on the multi-source raw data to obtain time synchronization data, and performing spatial coordinate conversion on the time synchronization data to obtain unified projection data; Performing data quality assessment on the unified projection data to obtain quality marked data; Performing multi-source data fusion on the quality-marked data to obtain preliminary fused data; Eliminate noise from the preliminary fusion data by using a Kalman filter algorithm to obtain noise reduction optimization data; Performing temporal and spatial scale consistency adjustment on the noise reduction and optimization data to obtain scale-uniform data; The scale-uniform data is discretized to obtain discretized time series data, and the discretized time series data is encoded with multi-dimensional features to obtain the standardized multi-dimensional time series precipitation data set.

3. The method for analyzing time series characteristics based on multidimensional data according to claim 1, characterized in that: The step of extracting spatiotemporal features and reducing dimension of the multi-dimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set includes: Calculating the fractal dimension of the multidimensional time series precipitation data set to obtain spatiotemporal complexity characteristics; Performing symbol dynamics encoding on the multi-dimensional time series precipitation data set to obtain symbol sequence features; Embedding the multidimensional time series precipitation data set by reconstructing the phase space to obtain phase space trajectory features; Performing Lyapunov exponent calculation on the phase space trajectory characteristics to obtain system stability characteristics; Performing recursive quantitative analysis on the multi-dimensional time series precipitation data set to obtain system periodic characteristics; Performing multi-scale entropy analysis on the multi-dimensional time series precipitation data set to obtain time series complexity characteristics; Performing time-frequency analysis on the multidimensional time series precipitation data set by Hilbert-Huang transform to obtain nonlinear time-frequency characteristics; Performing lagged mutual information analysis on the multidimensional time series precipitation dataset to obtain spatiotemporal dependency characteristics; Nonlinear dimensionality reduction is performed on the spatiotemporal complexity characteristics, the symbol sequence characteristics, the system stability characteristics, the system periodicity characteristics, the timing complexity characteristics, the nonlinear time-frequency characteristics and the spatiotemporal dependence characteristics to obtain the precipitation spatiotemporal feature vector set.

4. The method for analyzing time series characteristics based on multidimensional data according to claim 1, characterized in that: The multi-scale spatiotemporal pattern recognition and cluster analysis are performed on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results and an abnormal precipitation pattern list, including: Performing spatiotemporal sliding window segmentation on the precipitation spatiotemporal feature vector set to obtain a multi-scale feature segment set; Performing spatiotemporal pattern coding on the multi-scale feature segment set to obtain a pattern coding sequence; Performing frequent pattern mining on the pattern coding sequence to obtain a candidate spatiotemporal pattern set; Performing spatiotemporal association rule analysis on the candidate spatiotemporal pattern set to obtain a spatiotemporal association pattern; Clustering the spatiotemporal correlation patterns using a spectral clustering algorithm to obtain a preliminary clustering result; Performing fuzzy membership calculation on the preliminary clustering results to obtain soft clustering assignment; Performing density peak detection on the soft clustering assignment to obtain cluster centers; Redistributing the precipitation spatiotemporal feature vector set based on the cluster center to obtain a clustering result; Performing outlier detection on the clustering results to obtain a candidate set of abnormal precipitation patterns; quantifying the abnormal precipitation pattern candidate set by using an isolation forest algorithm to obtain the abnormal precipitation pattern list; The spatiotemporal correlation pattern, the clustering result and the abnormal precipitation pattern list are integrated to obtain the precipitation spatiotemporal pattern library.

5. The method for analyzing time series characteristics based on multidimensional data according to claim 1, characterized in that: The method of constructing a deep learning precipitation prediction model based on the precipitation spatiotemporal pattern library, the clustering results and the abnormal precipitation pattern list to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model includes: Embedding the precipitation spatiotemporal pattern library to obtain a pattern embedding matrix; Performing feature mapping on the clustering result to obtain a clustering feature tensor; Performing abnormal pattern encoding on the abnormal precipitation pattern list to obtain an abnormal feature vector; Performing multimodal feature fusion on the pattern embedding matrix, the clustering feature tensor and the abnormal feature vector to obtain fused feature representation data; Modeling the spatiotemporal dependency of the fused feature representation data through a graph convolutional network to obtain a spatiotemporal relationship graph; Performing attention mechanism calculation on the spatiotemporal relationship graph to obtain weighted spatiotemporal features; The weighted spatiotemporal features are sequence modeled by a bidirectional long short-term memory network to obtain a time series prediction result; A deconvolution operation is performed on the time series prediction results to obtain the end-to-end global atmospheric reanalysis precipitation fusion prediction model.

6. The method for analyzing time series characteristics based on multidimensional data according to claim 1, characterized in that: The dynamically updating and online learning of the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model includes: Perform data stream segmentation on real-time precipitation observation data to obtain time series data segments; Performing concept drift detection on the time series data segment to obtain a data distribution change indicator; Based on the data distribution change index, a parameter sensitivity analysis is performed on the global atmospheric reanalysis precipitation fusion prediction model to obtain a sensitive parameter set; Performing incremental learning on the sensitive parameter set to obtain an updated parameter vector; Performing model fusion on the updated parameter vector and the global atmospheric reanalysis precipitation fusion prediction model to obtain a temporary prediction model; Compressing the temporary prediction model to obtain a lightweight prediction model; Distributed training is performed on the lightweight prediction model through federated learning to obtain a multi-source collaborative optimization model; Performing model distillation on the multi-source collaborative optimization model to obtain a knowledge transfer model; Performing adversarial sample training on the knowledge transfer model to obtain a robustness enhanced model; Multi-task learning is performed on the robustness enhancement model to obtain the target precipitation spatiotemporal prediction model.

7. The method for analyzing time series characteristics based on multidimensional data according to claim 1, characterized in that: The multi-dimensional spatiotemporal analysis of the target precipitation spatiotemporal prediction model is performed to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy, including: Performing spatiotemporal decomposition on the output of the target precipitation spatiotemporal prediction model to obtain a multi-scale precipitation prediction result; Performing spatiotemporal autocorrelation analysis on the multi-scale precipitation prediction results to obtain precipitation pattern evolution characteristics; Performing topological data analysis on the precipitation pattern evolution characteristics to obtain a precipitation pattern topological structure; Based on the precipitation pattern topology, extreme value events are identified to obtain extreme precipitation risk assessment; Conduct scenario simulation on the extreme precipitation risk assessment to obtain multi-scenario precipitation impact prediction; Through complex network analysis, the system vulnerability assessment of the multi-scenario precipitation impact prediction is carried out to obtain key influencing factors; Causal inference is performed on the key influencing factors to obtain a precipitation system dynamics model; Based on the precipitation system dynamics model, a spatiotemporal knowledge graph is constructed to obtain a precipitation knowledge network; Perform interpretability analysis on the precipitation knowledge network to obtain the global atmospheric reanalysis precipitation fusion data analysis report; Based on the global atmospheric reanalysis precipitation fusion data analysis report, resource optimization allocation is performed to obtain the weather processing pipeline scheduling strategy.

8. A time series feature analysis system based on multidimensional data, used to execute the time series feature analysis method based on multidimensional data according to any one of claims 1 to 7, characterized in that: The time series feature analysis system based on multidimensional data includes: The acquisition module is used to collect and preprocess the multi-source raw data from weather observation stations, satellite remote sensing and ground observation stations to obtain a standardized multi-dimensional time series precipitation dataset; A dimension reduction module is used to extract spatiotemporal features and perform dimension reduction processing on the multidimensional time series precipitation data set to obtain a precipitation spatiotemporal feature vector set; A clustering module is used to perform multi-scale spatiotemporal pattern recognition and cluster analysis on the precipitation spatiotemporal feature vector set to obtain a precipitation spatiotemporal pattern library, clustering results and an abnormal precipitation pattern list; A construction module is used to construct a deep learning precipitation prediction model based on the precipitation spatiotemporal pattern library, the clustering results and the abnormal precipitation pattern list to obtain an end-to-end global atmospheric reanalysis precipitation fusion prediction model; An updating module is used to dynamically update and online learn the global atmospheric reanalysis precipitation fusion prediction model to obtain a target precipitation spatiotemporal prediction model; The analysis module is used to perform multi-dimensional spatiotemporal analysis on the target precipitation spatiotemporal prediction model to obtain a global atmospheric reanalysis precipitation fusion data analysis report and a weather processing pipeline scheduling strategy.

Images