A weather element forecast generation method and system based on multi-source information fusion

Abstract

The application provides a meteorological element forecast generation method and system based on multi-source information fusion, and relates to the technical field of meteorological forecast. The application collects grid data of third-party meteorological software, converts the grid data into binary meteorological labels, and delimits a historical and backtracking time window. Mutual information values of a meteorological prediction target and a candidate grid feature sequence are calculated, effective grid points are screened, time sequence samples are constructed based on the effective grid points, a frequent time sequence pattern mining algorithm with time tolerance is used, meteorological evolution path sets are extracted, path confidence is calculated in combination with historical frequency, similarity between a meteorological event set and the meteorological evolution path set is calculated, and a synergistic gain term is constructed by multiplying the two. The confidence, maximum similarity and synergistic gain term output by the meteorological software are input into a trained multi-classification logistic regression model, and a target with the maximum probability is selected as a final prediction result, so as to make up for the prediction deviation and scene adaptability limitation of a single third-party meteorological software.

Description

Technical Field

[0001] This invention relates to the field of meteorological element forecasting technology, specifically to a meteorological element forecasting generation method and system based on multi-source information fusion. Background Technology

[0002] In recent years, with the development of numerical weather prediction technology, a variety of meteorological fusion analysis software has emerged, providing a gridded meteorological analysis field covering the globe, and has been widely used in weather diagnosis and forecasting operations.

[0003] Currently, various fusion analysis software programs are widely used in mainstream meteorological operations both domestically and internationally. These include ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF) and software from the U.S. Global Forecast System (GRS), providing global gridded meteorological analysis fields and achieving good results in large-scale weather system forecasting and climate monitoring, demonstrating a certain forecasting capability. However, for specific small areas, such as chemical plant areas, there are still shortcomings in forecasting meteorological elements. The spatial resolution of these software programs is coarse, and the representativeness of sparsely observed areas is poor. Most of the data corresponds to a ground level of 25–50 kilometers, which is much larger than the scale of small areas like chemical plants. In areas with insufficient meteorological observation stations, the grid data from fusion analysis software mainly relies on satellite remote sensing and numerical model extrapolation, resulting in poor local representativeness and large errors. This makes it difficult to directly meet the needs of accurate forecasting for small-scale areas. Consequently, when relying on a single software program, it is impossible to adaptively learn the reliability weights of each software program under different regions and weather types, leading to insufficient ability of forecast models to capture key precursor signals before severe weather events.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for generating meteorological element forecasts based on multi-source information fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for generating meteorological element forecasts based on multi-source information fusion, comprising the following steps: S1. For each third-party meteorological software, collect meteorological tags of candidate grid points, set historical time windows, and several backtracking time windows with a preset time interval from the historical time windows, and select the meteorological forecast targets to be extrapolated from the collected meteorological tags. S2. Based on the binarized meteorological labels of the historical time window of the area to be analyzed, construct the target time series of each meteorological forecast target, construct the feature variable series based on the meteorological labels of the backtracking time window of each candidate grid point, calculate the mutual information value between the target time series and each feature variable series, and screen out the effective grid points whose correlation reaches the preset threshold. S3. Based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target, time series samples are constructed. A frequent time series pattern mining algorithm with time tolerance is used to mine multiple sets of meteorological evolution paths. S4. Combining the historical occurrence frequency and historical time window binarized labels of each meteorological evolution path set, calculate the confidence of each meteorological evolution path set. Construct a meteorological event set using the backtracking time window meteorological labels, calculate the similarity between the meteorological event set and the meteorological evolution path set, and use the product of the maximum similarity and the confidence of the corresponding meteorological evolution path set as the collaborative gain term. S5. At the current forecast time, input the confidence scores, the corresponding calculated maximum similarity scores, and the collaborative gain terms from the outputs of each meteorological software into the trained regression model. Output the probability of weather occurrence for each meteorological prediction target. Select the meteorological prediction target with the maximum probability as the final meteorological forecast result for the area to be analyzed.

[0007] Furthermore, based on the binarized meteorological labels of the historical time windows of the area to be analyzed, the target time series of each meteorological forecast objective is constructed. The specific method is as follows: The binarized meteorological label is the binary determination result of whether various meteorological events have occurred in a predefined time window for each candidate grid point in the area to be analyzed. For any weather forecast target, several retrospective time windows are selected from historical data. The binarized weather labels of each candidate grid point in the area to be analyzed corresponding to all historical time windows are arranged in chronological order to construct the target time series corresponding to the weather forecast target.

[0008] Furthermore, a sequence of feature variables is constructed based on the meteorological labels of the backtracking time windows of each candidate grid point. The specific method is as follows: For any candidate grid point, each backtracking time window with a preset interval from the historical time window is used as a unit, and the meteorological labels within the backtracking time window are extracted as a set of variable data. All variable data corresponding to the backtracking time windows are arranged in chronological order consistent with the target time series to construct the feature variable sequence corresponding to the candidate grid point. The feature variable sequence has the same length as the target time series, and each set of variable data corresponds one-to-one with each binarized meteorological label in the target time series and is aligned with the time reference.

[0009] Furthermore, the specific method for calculating the mutual information values ​​between the target time series and each feature variable series, and filtering out valid grid points whose correlation reaches a preset threshold, is as follows: For any weather forecast target, the mutual information value between the time series sequence of the target and the feature variable sequence corresponding to each candidate grid point is calculated. The feature variable sequence with a mutual information value higher than the preset information threshold is retained, and the candidate grid point corresponding to the feature variable sequence is determined as the effective grid point of the weather forecast target.

[0010] Furthermore, time-series samples are constructed based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target. The specific method is as follows: For any weather forecast target, select any set of retrospective time windows with a preset time interval from the historical time window, extract the weather tags corresponding to each effective grid point in the set of retrospective time windows, and combine all effective grid points and their weather tags in the same retrospective time window into a set of weather events. Arrange the sets of meteorological events corresponding to all backtracking time windows in chronological order to form a time series sample. By traversing all backtracking time windows, multiple sets of time series samples corresponding to the meteorological forecast target are obtained.

[0011] Furthermore, a frequent time-series pattern mining algorithm with time tolerance is used to mine multiple sets of meteorological evolution paths. The specific method is as follows: A minimum support threshold is preset. For any weather forecast target, all time series samples corresponding to the weather forecast target are used as input to perform frequent time series pattern mining with time tolerance constraints. The mined frequent time series patterns that meet the minimum support threshold are defined as the set of weather evolution paths, and finally multiple sets of weather evolution paths corresponding to the weather forecast target are obtained. The constraints for the mining process are as follows: The smallest unit in a single path set is an event item, which consists of a single valid grid point and its corresponding meteorological label. The backtracking time window corresponding to a later event item in the path set is no earlier than the backtracking time window corresponding to the previous event item, maintaining the temporal order. The same valid grid point appears at most once in the same path set. The time interval between the backtracking time windows corresponding to adjacent event items in the path set is within a preset time tolerance range.

[0012] Furthermore, by combining the historical occurrence frequency and the binarized labels of historical time windows of each meteorological evolution path set, the confidence level of each meteorological evolution path set is calculated. The specific method is as follows: For any set of meteorological evolution paths, the total number of time series samples containing that path set in all time series samples corresponding to the meteorological forecast target is counted, and this number is taken as the path occurrence frequency. In the time series samples containing the path set, the number of time series samples in which the binarized meteorological label of the meteorological forecast target is in the occurrence state within the corresponding historical time window is counted, and this number is used as the target matching frequency. Calculate the ratio of the target matching frequency to the path occurrence frequency, and determine this ratio as the confidence level of the meteorological evolution path set.

[0013] Furthermore, a set of meteorological events is constructed using backtracking time window meteorological labels. The similarity between the meteorological event set and the path set is calculated, and the set of meteorological evolution paths corresponding to the highest similarity is selected. The specific method is as follows: Extract the meteorological tags corresponding to all valid grid points within each retrospective time window at the current forecast time, and integrate them to form the current meteorological event set; calculate the matching similarity between this meteorological event set and each set of meteorological evolution paths, and quantify the degree of overlap between the valid grid points, meteorological tags, and temporal arrangement contained in the two; traverse the similarity values ​​corresponding to all meteorological evolution path sets, filter out the maximum similarity, and lock the meteorological evolution path set corresponding to the maximum similarity.

[0014] Additionally, a meteorological element forecast generation system based on multi-source information fusion is provided. This system is used to execute any of the above-described meteorological element forecast generation methods based on multi-source information fusion, including: The data extraction module is used to collect meteorological labels of candidate grid points for various third-party meteorological software, set historical time windows, and several backtracking time windows with a preset time interval from the historical time windows, and select the meteorological forecast targets to be extrapolated from the collected meteorological labels. The grid point filtering module is used to construct target time series of each meteorological forecast target based on the binarized meteorological labels of the historical time window of the area to be analyzed, construct feature variable series based on the meteorological labels of the backtracking time window of each candidate grid point, calculate the mutual information value between the target time series and each feature variable series, and filter out effective grid points whose correlation reaches a preset threshold. The path set filtering module is used to construct time series samples based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target. It uses a frequent time series pattern mining algorithm with time tolerance to mine multiple sets of meteorological evolution paths. The feature calculation module is used to combine the historical occurrence frequency of each meteorological evolution path set with the binarized labels of the historical time window to calculate the confidence of each meteorological evolution path set. It constructs a meteorological event set using the meteorological labels of the backtracking time window, calculates the similarity between the meteorological event set and the meteorological evolution path set, and uses the product of the maximum similarity and the confidence of the corresponding meteorological evolution path set as the collaborative gain term. The weather forecast module is used to input the confidence scores, the corresponding calculated maximum similarity, and the collaborative gain terms from the outputs of various weather software into the trained regression model at the current forecast time. It outputs the probability of weather occurrence for each weather forecast target and selects the weather forecast target with the maximum probability as the final weather forecast result for the area to be analyzed.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention extracts meteorological tags from candidate grid points, calculates mutual information, and selects effective grid points with a correlation reaching a preset threshold. It constructs time-series samples based on the effective grid points corresponding to each prediction target and the meteorological tags of the retrospective time window. It mines path sets using a frequent time-series pattern with time tolerance. When it is necessary to predict meteorological results for a preset duration, it provides a more accurate prediction data benchmark for subsequent weather forecasts by analyzing the meteorological tags of effective grid points in different retrospective time windows. The present invention also calculates the confidence level of each set of meteorological evolution paths, calculates the similarity between the set of meteorological events and the set of meteorological evolution paths, and uses the product of the maximum similarity and the confidence level of the corresponding set of meteorological evolution paths as a collaborative gain term. The confidence scores, the corresponding calculated maximum similarity scores, and the collaborative gain terms output by each meteorological software are input into the trained regression model. The model outputs the probability of weather occurrence for each predicted target. The predicted target corresponding to the maximum probability is selected as the final meteorological forecast result for the area to be analyzed. This compensates for the weather forecast bias and scenario adaptability limitations of a single meteorological software under different weather systems, thereby improving the accuracy of the final weather forecast probability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the mutual information value calculation results of the present invention; Figure 3 This is a comparison chart of the probability of weather occurrence according to the present invention; Figure 4 This is a block diagram of the overall system modules of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] Example: Please see Figures 1 to 3 The present invention provides a technical solution: A method for generating meteorological element forecasts based on multi-source information fusion, comprising the following steps: S1. For each third-party meteorological software, collect meteorological tags of candidate grid points, set historical time windows, and several backtracking time windows with a preset time interval from the historical time windows, and select the meteorological forecast targets to be extrapolated from the collected meteorological tags. For the region to be analyzed, first identify multiple third-party meteorological data platforms that provide meteorological data, including but not limited to the European Centre for Medium-Range Weather Forecasts, the Global Weather Forecast System, or domestic commercial meteorological APIs.

[0020] Due to spatial resolution limitations, meteorological data for any location needs to be approximated using its corresponding grid. A common meteorological grid is pre-defined under a unified geographic coordinate system, and meteorological fields output by various third-party meteorological data platforms are mapped to this common meteorological grid based on the nearest neighbor or maximum coverage principle. To avoid ambiguity caused by overlapping meteorological data, the mapped meteorological data is converted into a set of mutually exclusive meteorological labels, such as sunny, cloudy, overcast, and different levels of precipitation. This ensures that any candidate grid point is uniquely assigned to one of these categories within any fixed time window. Finally, the meteorological forecast target to be extrapolated is selected from all collected meteorological labels.

[0021] Given that weather systems have advection and movement characteristics, the weather signals that appear in the surrounding grids in the early stages often have indicative significance for the future weather of the area to be analyzed. Therefore, a meteorological influence area is formed by expanding outward from the area to be analyzed, and all grid points within this area are identified as candidate grid points, that is, grid points that may be moved towards the area to be analyzed by the meteorological system. For example, a range with a radius of 200 km centered on the geographical center of the area to be analyzed is used, and all grid points within this range are included in the candidate grid point set.

[0022] Constructing time-series samples: Historical time windows are used as the prediction periods in model training, and the meteorological labels that actually occur in the area to be analyzed within the historical time window are recorded. The preset duration refers to the time interval between the end time of the retrospective time window and the start time of the historical time window, which represents the lead time of the weather forecast. For example, if the preset duration is set to 72 hours, it means that the trained model can use historical input features to predict the meteorological labels of the area to be analyzed 3 days in advance. The number of retrospective time windows represents the duration of historical data that needs to be observed. It is preferred to set it to 3-5 to avoid the inability to capture the movement speed and evolution trend of the meteorological system if the number is too small, or the introduction of invalid interference information and the increase of computational complexity for subsequent frequent time-series pattern mining if the number is too large. Through the above division of time windows, the meteorological prediction target corresponding to the area to be analyzed within the historical time window is used as the result label of the training sample, and the meteorological labels of each candidate grid point within each retrospective time window are used as the input features for inferring the result label. The duration of all time windows is kept consistent, such as being set to 6 hours.

[0023] S2. Based on the binarized meteorological labels of the historical time window of the area to be analyzed, construct the target time series of each meteorological forecast target, construct the feature variable series based on the meteorological labels of the backtracking time window of each candidate grid point, calculate the mutual information value between the target time series and each feature variable series, and screen out the effective grid points whose correlation reaches the preset threshold. The data formats corresponding to different meteorological forecast targets are not consistent. For example, sunny, cloudy, and rainy weather are categorical data, while precipitation, temperature, and wind speed are continuous data. Therefore, for any meteorological forecast target, the actual meteorological labels of the area to be analyzed within each historical time window are binarized to generate binarized meteorological labels. The binarized meteorological labels are the binary judgment results of whether various meteorological events have occurred in the candidate grid points within the historical time window. They are used to filter numerical fluctuations and make the model focus on the characteristics of whether the meteorological label has occurred. This not only allows for the analysis of the correlation between different meteorological targets and candidate grid points, but also facilitates the subsequent model to output the occurrence probability of each meteorological forecast target. For any weather forecast target, several retrospective time windows are selected from historical data. The binarized weather labels of each candidate grid point in the area to be analyzed corresponding to all historical time windows are arranged in chronological order to construct the target time series corresponding to the weather forecast target.

[0024] For any candidate grid point, a backtracking time window with a preset interval from the historical time window is used as a unit to extract meteorological labels within the backtracking time window as a set of variable data. All variable data corresponding to the backtracking time windows are arranged in chronological order consistent with the target time series to construct the feature variable sequence corresponding to the candidate grid point. The feature variable sequence has the same length as the target time series, and each set of variable data corresponds one-to-one with each binarized meteorological label in the target time series and is aligned with the time reference.

[0025] In weather and meteorological systems, processes such as cloud movement and frontal evolution exhibit strong nonlinear dynamic characteristics. Mutual information, based on information entropy theory, can sensitively capture complex nonlinear dependencies between variables. Therefore, for any meteorological forecast target, the mutual information value between the target time series and the characteristic variable series corresponding to each candidate grid point is calculated. The mutual information value is used to quantify the amount of information about the target time series contained in the characteristic variable series, such as... Figure 2 As shown, the mutual information value between a certain weather forecast target and 5 candidate grid points is calculated. The feature variable sequence with mutual information value higher than the preset information threshold is retained, and the candidate grid points corresponding to the feature variable sequence are determined as the effective grid points of the weather forecast target. Since the mutual information value itself is affected by the number of samples and data distribution, an adaptive threshold dynamically calculated based on the data distribution should be used as the information threshold. The mutual information values ​​corresponding to each candidate grid point are arranged in descending order to construct a mutual information decay curve. The point of maximum curvature of the decay curve is calculated as the inflection point, and the mutual information value corresponding to the inflection point is determined as the preset threshold. For example, when the preset threshold is 0.40, grid points 2, 3, and 5 with mutual information values ​​higher than the preset threshold are retained, and the three are determined as the effective grid points of the weather forecast target. Different third-party meteorological software may use different data sources, numerical forecasting models, and spatial resolutions, and their prediction bias and response lead time may also differ. Therefore, the above mutual information calculation and screening operations are performed independently for each third-party meteorological software. This is to evaluate the spatial correlation and meteorological data validity of each third-party meteorological software at each candidate grid point when dealing with the data differences among the software, thereby fully preserving the unique data characteristics of each software. At the same time, by screening effective grid points through mutual information, grid points with low correlation to other meteorological prediction targets can be excluded, reducing the number of event items in subsequent time series samples. This reduces the combined scale of frequent time series pattern mining and avoids diluting the real meteorological evolution path with a large number of irrelevant meteorological labels.

[0026] S3. Based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target, time series samples are constructed. A frequent time series pattern mining algorithm with time tolerance is used to mine multiple sets of meteorological evolution paths. For any weather forecast target, select any set of retrospective time windows with a preset time interval from the historical time window, extract the weather tags corresponding to each effective grid point in the set of retrospective time windows, and combine all effective grid points and their weather tags in the same retrospective time window into a set of weather events. All meteorological events corresponding to the backtracking time windows are arranged in chronological order to form a time series sample. By traversing all backtracking time windows, multiple sets of time series samples corresponding to the meteorological prediction target are obtained. These multiple sets of time series samples objectively record the meteorological labels of each effective grid point in different backtracking time windows when a specific meteorological prediction target occurs in the area to be analyzed. The essence of performing frequent time series pattern mining is to find subsequences with strong statistical correlation characteristics in massive historical data. The high-frequency time series event combinations that meet the support threshold constraint reflect the transfer trajectory of meteorological state in the spatial dimension and characterize the chronological order of meteorological evolution in the temporal dimension. A minimum support threshold is preset. For any weather forecast target, all time series samples corresponding to the weather forecast target are used as input to perform frequent time series pattern mining with time tolerance constraints. The minimum subsequence length of the mined pattern must be greater than 1. The mined frequent time series patterns that meet the minimum support threshold are defined as the set of weather evolution paths. Finally, multiple sets of weather evolution paths corresponding to the weather forecast target are obtained. If the minimum support threshold is too high, it may miss extreme weather evolution paths that occur less frequently but have strong predictive power. Therefore, the minimum support threshold is set to a range of 0.05 to 0.15. Traditional frequent sequence mining algorithms usually require strict time alignment. However, in the real natural atmospheric environment, the advection and movement speed of precipitation clouds and cold fronts between spatial grids are not constant. Their movement process is affected by complex dynamic factors such as changes in atmospheric circulation, local topographic undulations, and differences in surface thermodynamics, which inevitably lead to deceleration or lag in propagation. By introducing time tolerance and setting the time tolerance to an integer multiple of the time window length, the time interval between adjacent events is allowed to fluctuate within a reasonable range, while maintaining the order between events in the path. This improves the ability to summarize similar meteorological evolution processes. For example, the time tolerance is set to the length of the time window, allowing the time interval between adjacent events to fluctuate within a time window. The constraints for the mining process are as follows: The smallest unit in a single path set is an event item, which consists of a single valid grid point and its corresponding meteorological tag. Based on the unidirectional nature of cloud and front evolution and the sequential order of meteorological precursor signals, the retrospective time window corresponding to a later event item in the path set is no earlier than the retrospective time window corresponding to a previous event item, maintaining the temporal sequence. The same valid grid point appears at most once in the same path set. By restricting the recurrence of single grid points, the constraint algorithm focuses attention on the movement trajectory of meteorological tags across grids. The time interval between the retrospective time windows corresponding to adjacent event items in the path set is within a preset time tolerance range, solving the lag characteristics in meteorological propagation.

[0027] S4. Combining the historical occurrence frequency and historical time window binarized labels of each meteorological evolution path set, calculate the confidence of each meteorological evolution path set. Construct a meteorological event set using the backtracking time window meteorological labels, calculate the similarity between the meteorological event set and the meteorological evolution path set, and use the product of the maximum similarity and the confidence of the corresponding meteorological evolution path set as the collaborative gain term. For any set of meteorological evolution paths, the total number of time series samples containing this path set in all time series samples corresponding to the meteorological forecast target is counted, and this number is taken as the path occurrence frequency. In the time series samples containing this path set, the number of time series samples in which the binarized meteorological label of the meteorological forecast target is in the occurrence state within the corresponding historical time window is counted, and this number is used as the target matching frequency. The ratio of the target matching frequency to the path occurrence frequency is calculated, and this ratio is determined as the confidence level of the meteorological evolution path set. The path mined by minimum support only represents that the meteorological sequence has occurred frequently in history, but does not indicate that the target meteorological event will definitely occur. The confidence level is used to quantify the correlation between the corresponding meteorological prediction target in the area to be analyzed under the path. All historical event items in a set of meteorological evolution paths are extracted to form a target sequence. Simultaneously, the backtracking time window is traversed to extract all event items, forming the current real-time sequence. Event items from the target sequence are retrieved one by one within the real-time sequence, and the number of successfully retrieved event items is counted. The ratio of this number to the total number of event items in the target sequence is defined as the event item hit rate. For the successfully retrieved event items, their actual occurrence time sequence in the current real-time sequence is extracted to construct the current occurrence subsequence. Simultaneously, the original time sequence of these event items in the historical meteorological evolution path set is extracted to construct the historical occurrence subsequence. The longest common subsequence algorithm is used to compare the temporal arrangement consistency between the current occurrence subsequence and the historical occurrence subsequence. The ratio of the length of the longest common subsequence to the total number of event items in the historical occurrence subsequence is used as the temporal sequence fit. Finally, the product of the event item hit rate and the temporal sequence fit is used as the similarity between the current real-time sequence and a single set of meteorological evolution paths. This product form is subject to a dual necessary condition constraint: the similarity of meteorological paths requires the simultaneous satisfaction of two judgment dimensions. First, meteorological events in historical paths can be effectively detected in the current sequence. Second, the order of occurrence of detected events conforms to the objective evolution law of the movement, generation and dissipation of meteorological systems.

[0028] The product calculation method has strong screening characteristics: if the event hit rate approaches 0, or the temporal sequence fit approaches 0, the similarity will also approach 0, and it can be directly determined that the two meteorological evolution paths do not have reference similarity; only when the detected event proportion is high and the event occurrence sequence is highly consistent with the historical meteorological evolution logic, the similarity value approaches 1, which means that the current real sequence is highly matched with the historical meteorological evolution path.

[0029] The similarity values ​​between the current set of meteorological events and all sets of meteorological evolution paths are traversed, and the maximum similarity is selected. The product of the maximum similarity and the confidence of the corresponding set of meteorological evolution paths is used as a collaborative gain term. The value of the collaborative gain term will only be increased when the meteorological data highly matches a certain path in history, and the path has indeed historically had a high probability of causing the meteorological prediction target to occur. This enables adaptive correction of the prediction bias of third-party meteorological software in subsequent steps.

[0030] S5. At the current forecast time, input the confidence scores, the corresponding calculated maximum similarity and collaborative gain terms of each meteorological software into the trained regression model, output the probability of weather occurrence for each meteorological prediction target, and select the meteorological prediction target corresponding to the maximum probability as the final meteorological forecast result for the area to be analyzed. In the historical sample construction stage, the confidence level, maximum similarity, and collaborative gain term output by the third-party meteorological software for each meteorological forecast target are concatenated into a one-dimensional feature vector, which is used as the input feature of the regression model; and the actual meteorological categories that occurred in the area to be analyzed during the corresponding historical period are used as supervision labels to complete the construction of the training dataset.

[0031] The regression model employs a multi-class logistic regression model, using multi-class cross-entropy as the loss function, and incorporating an L2 regularization term to suppress overfitting. The optimal model parameters are solved by minimizing the total loss function with L2 regularization constraints. These parameters include: the software confidence weights, maximum similarity weights, and collaborative gain weights for each meteorological prediction target, as well as a bias term for that meteorological prediction target. This bias term characterizes the inherent probability of occurrence of this type of meteorological event without introducing any input features.

[0032] The model training process is as follows: First, the linear score of each meteorological forecast target is calculated, and all linear scores are converted into corresponding meteorological occurrence probabilities using a normalized exponential function to ensure that the sum of the meteorological occurrence probabilities of each meteorological forecast target is 1 at the same historical forecast time. The mapped predicted probability distribution is compared with the historical real meteorological labels, and the deviation between the prediction results and the real categories is quantified using multi-class cross-entropy loss. The smaller the cross-entropy loss value, the higher the predicted probability assigned to the real meteorological target. Based on this, an L2 regularization term is used to apply a coefficient penalty constraint to all feature weights, and the sum of the cross-entropy loss and the L2 regularization penalty term is used as the objective function for minimization. After iterative convergence, the optimal feature weights and bias terms for each group of meteorological forecast targets are obtained. After the model training is completed and converged, a feature vector consisting of confidence, maximum similarity, and collaborative gain terms can be input into the model in real time, and the occurrence probability of each meteorological forecast target can be directly output to complete the rolling meteorological discrimination.

[0033] During the model inference phase, for the current forecast time, real-time meteorological labels corresponding to all valid grid points within each backtracking time window are extracted and integrated to form a current meteorological event set. The matching similarity between this current meteorological event set and each set of meteorological evolution paths is calculated, and the similarity is quantified based on the degree of overlap of the valid grid points, meteorological labels, and temporal arrangement contained in the two sets. The similarity values ​​corresponding to all meteorological evolution path sets are traversed, and the maximum similarity at the current time is selected. The meteorological evolution path set and the collaborative gain term corresponding to the maximum similarity are then locked. Finally, the real-time confidence scores, the currently calculated maximum similarity, and the corresponding collaborative gain terms output by each meteorological software for each meteorological forecast target are concatenated into a real-time feature vector. This vector is then input into the trained multi-class logistic regression model to output the probability of each meteorological forecast target occurring at the current forecast time. The meteorological forecast target with the highest probability is selected as the final meteorological forecast result for the area to be analyzed. Figure 3 As shown in the figure, at a certain moment, the process of multi-source information fusion and probability calculation for five mutually exclusive weather forecast targets is performed using a trained multi-class logistic regression model. First, the initial confidence scores output by Software 1, Software 2, and Software 3 for each forecast target are obtained. Then, these confidence scores are concatenated with the corresponding calculated maximum similarity and collaborative gain terms to form a feature vector, which is input into the regression model. The final weather occurrence probability of each weather forecast target is output. Finally, by iterating and comparing the output probabilities, the weather occurrence probability corresponding to moderate rain is the absolute maximum among all forecast targets. Therefore, moderate rain is selected as the final weather forecast result for the area to be analyzed.

[0034] Please see Figure 4 The present invention also provides a meteorological element forecast generation system based on multi-source information fusion, the system being used to execute the meteorological element forecast generation method based on multi-source information fusion described in any of the above claims, comprising: The data extraction module is used to collect meteorological labels of candidate grid points for various third-party meteorological software, set historical time windows, and several backtracking time windows with a preset time interval from the historical time windows, and select the meteorological forecast targets to be extrapolated from the collected meteorological labels. The grid point filtering module is used to construct target time series of each meteorological forecast target based on the binarized meteorological labels of the historical time window of the area to be analyzed, construct feature variable series based on the meteorological labels of the backtracking time window of each candidate grid point, calculate the mutual information value between the target time series and each feature variable series, and filter out effective grid points whose correlation reaches a preset threshold. The path set filtering module is used to construct time series samples based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target. It uses a frequent time series pattern mining algorithm with time tolerance to mine multiple sets of meteorological evolution paths. The feature calculation module is used to combine the historical occurrence frequency of each meteorological evolution path set with the binarized labels of the historical time window to calculate the confidence of each meteorological evolution path set. It constructs a meteorological event set using the meteorological labels of the backtracking time window, calculates the similarity between the meteorological event set and the meteorological evolution path set, and uses the product of the maximum similarity and the confidence of the corresponding meteorological evolution path set as the collaborative gain term. The weather forecast module is used to input the confidence scores, the corresponding calculated maximum similarity, and the collaborative gain terms from the outputs of various weather software into the trained regression model at the current forecast time. It outputs the probability of weather occurrence for each weather forecast target and selects the weather forecast target with the maximum probability as the final weather forecast result for the area to be analyzed.

[0035] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0036] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0037] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for generating meteorological element forecasts based on multi-source information fusion, characterized in that, The specific steps include: S1. For each third-party meteorological software, collect meteorological tags of candidate grid points, set historical time windows, and several backtracking time windows with a preset time interval from the historical time windows, and select the meteorological forecast targets to be extrapolated from the collected meteorological tags. S2. Based on the binarized meteorological labels of the historical time window of the area to be analyzed, construct the target time series of each meteorological forecast target, construct the feature variable series based on the meteorological labels of the backtracking time window of each candidate grid point, calculate the mutual information value between the target time series and each feature variable series, and screen out the effective grid points whose correlation reaches the preset threshold. S3. Based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target, time series samples are constructed. A frequent time series pattern mining algorithm with time tolerance is used to mine multiple sets of meteorological evolution paths. S4. Combining the historical occurrence frequency and historical time window binarized labels of each meteorological evolution path set, calculate the confidence of each meteorological evolution path set. Construct a meteorological event set using the backtracking time window meteorological labels, calculate the similarity between the meteorological event set and the meteorological evolution path set, and use the product of the maximum similarity and the confidence of the corresponding meteorological evolution path set as the collaborative gain term. S5. At the current forecast time, input the confidence scores, the corresponding calculated maximum similarity scores, and the collaborative gain terms from the outputs of each meteorological software into the trained regression model. Output the probability of weather occurrence for each meteorological prediction target. Select the meteorological prediction target with the maximum probability as the final meteorological forecast result for the area to be analyzed.

2. The meteorological element forecast generation method based on multi-source information fusion according to claim 1, characterized in that: Based on the binarized meteorological labels of the historical time windows of the area to be analyzed, the target time series of each meteorological forecast objective is constructed. The specific method is as follows: The binarized meteorological label is the binary determination result of whether various meteorological events have occurred in a predefined time window for each candidate grid point in the area to be analyzed. For any weather forecast target, several retrospective time windows are selected from historical data. The binarized weather labels of each candidate grid point in the area to be analyzed corresponding to all historical time windows are arranged in chronological order to construct the target time series corresponding to the weather forecast target.

3. The meteorological element forecast generation method based on multi-source information fusion according to claim 1, characterized in that: The feature variable sequence is constructed based on the meteorological labels of each candidate grid point's backtracking time window. The specific method is as follows: For any candidate grid point, each set of retrospective time windows with a preset interval from the historical time window is used as a unit, and the meteorological labels within the retrospective time window are extracted as a set of variable data; all the variable data corresponding to the retrospective time windows are arranged in chronological order consistent with the target time series to construct the feature variable sequence corresponding to the candidate grid point. The feature variable sequence is of equal length to the target time series sequence, and each group of variable data corresponds one-to-one with each binarized meteorological label in the target time series sequence and is aligned with the time reference.

4. The meteorological element forecast generation method based on multi-source information fusion according to claim 1, characterized in that: The specific method for calculating the mutual information values ​​between the target time series and each feature variable series, and filtering out valid grid points whose correlation reaches a preset threshold, is as follows: For any weather forecast target, the mutual information value between the time series sequence of the target and the feature variable sequence corresponding to each candidate grid point is calculated. The feature variable sequence with a mutual information value higher than the preset information threshold is retained, and the candidate grid point corresponding to the feature variable sequence is determined as the effective grid point of the weather forecast target.

5. The meteorological element forecast generation method based on multi-source information fusion according to claim 4, characterized in that: Time-series samples are constructed based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target. The specific method is as follows: For any weather forecast target, select any set of retrospective time windows with a preset time interval from the historical time window, extract the weather tags corresponding to each effective grid point in the set of retrospective time windows, and combine all effective grid points and their weather tags in the same retrospective time window into a set of weather events. Arrange the sets of meteorological events corresponding to all backtracking time windows in chronological order to form a time series sample. By traversing all backtracking time windows, multiple sets of time series samples corresponding to the meteorological forecast target are obtained.

6. The meteorological element forecast generation method based on multi-source information fusion according to claim 1, characterized in that: A frequent time series pattern mining algorithm with time tolerance is used to mine multiple sets of meteorological evolution paths. The specific method is as follows: A minimum support threshold is preset. For any weather forecast target, all time series samples corresponding to the weather forecast target are used as input to perform frequent time series pattern mining with time tolerance constraints. The mined frequent time series patterns that meet the minimum support threshold are defined as the set of weather evolution paths, and finally multiple sets of weather evolution paths corresponding to the weather forecast target are obtained. The constraints for the mining process are as follows: The smallest unit in a single path set is an event item, which consists of a single valid grid point and its corresponding meteorological label. The backtracking time window corresponding to a later event item in the path set is no earlier than the backtracking time window corresponding to the previous event item, maintaining the temporal order. The same valid grid point appears at most once in the same path set. The time interval between the backtracking time windows corresponding to adjacent event items in the path set is within a preset time tolerance range.

7. The meteorological element forecast generation method based on multi-source information fusion according to claim 1, characterized in that: By combining the historical occurrence frequency and the binarized labels of historical time windows of each meteorological evolution path set, the confidence level of each meteorological evolution path set is calculated. The specific method is as follows: For any set of meteorological evolution paths, the total number of time series samples containing that path set in all time series samples corresponding to the meteorological forecast target is counted, and this number is taken as the path occurrence frequency. In the time series samples containing the path set, the number of time series samples in which the binarized meteorological label of the meteorological forecast target is in the occurrence state within the corresponding historical time window is counted, and this number is used as the target matching frequency. Calculate the ratio of the target matching frequency to the path occurrence frequency, and determine this ratio as the confidence level of the meteorological evolution path set.

8. The meteorological element forecast generation method based on multi-source information fusion according to claim 1, characterized in that: A set of meteorological events is constructed using backtracking time window meteorological labels. The similarity between the meteorological event set and the path set is calculated, and the set of meteorological evolution paths corresponding to the highest similarity is selected. The specific method is as follows: Extract the meteorological labels corresponding to all valid grid points within each retrospective time window at the current forecast time, and integrate them to form the current meteorological event set; calculate the matching similarity between this meteorological event set and each set of meteorological evolution paths, and quantify the degree of overlap between the valid grid points, meteorological labels, and temporal arrangement contained in the two; Traverse all similarity values ​​corresponding to the set of meteorological evolution paths, filter to obtain the maximum similarity, and lock the set of meteorological evolution paths corresponding to the maximum similarity.

9. A meteorological element forecast generation system based on multi-source information fusion, characterized in that: The system is used to execute the meteorological element forecast generation method based on multi-source information fusion as described in any one of claims 1-8: The data extraction module is used to collect meteorological labels of candidate grid points for various third-party meteorological software, set historical time windows, and several backtracking time windows with a preset time interval from the historical time windows, and select the meteorological forecast targets to be extrapolated from the collected meteorological labels. The grid point filtering module is used to construct target time series of each meteorological forecast target based on the binarized meteorological labels of the historical time window of the area to be analyzed, construct feature variable series based on the meteorological labels of the backtracking time window of each candidate grid point, calculate the mutual information value between the target time series and each feature variable series, and filter out effective grid points whose correlation reaches a preset threshold. The path set filtering module is used to construct time series samples based on the effective grid points and backtracking time window meteorological labels corresponding to each meteorological forecast target. It uses a frequent time series pattern mining algorithm with time tolerance to mine multiple sets of meteorological evolution paths. The feature calculation module is used to combine the historical occurrence frequency of each meteorological evolution path set with the binarized labels of the historical time window to calculate the confidence of each meteorological evolution path set. It constructs a meteorological event set using the meteorological labels of the backtracking time window, calculates the similarity between the meteorological event set and the meteorological evolution path set, and uses the product of the maximum similarity and the confidence of the corresponding meteorological evolution path set as the collaborative gain term. The weather forecast module is used to input the confidence scores, the corresponding calculated maximum similarity, and the collaborative gain terms from the outputs of various weather software into the trained regression model at the current forecast time. It outputs the probability of weather occurrence for each weather forecast target and selects the weather forecast target with the maximum probability as the final weather forecast result for the area to be analyzed.

Images