Predicting device and predicting method
The integration of spatial and temporal feature extraction in a machine learning-based prediction model addresses the limitations of localized data use in river flow rate prediction, resulting in improved accuracy through the use of comprehensive mesh data.
Patent Information
- Application Number
- JP2023083800
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-22
- Publication Date
- 2024-12-04
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing river flow rate prediction models are limited by the use of localized rainfall data, failing to incorporate the comprehensive rainfall situation of the entire catchment area, which hinders accurate prediction.
A prediction device and method that utilizes mesh data through machine learning, incorporating a spatial and temporal feature extraction process to construct a flow rate prediction model, enabling the integration of large-scale mesh data for improved accuracy.
Enhances prediction accuracy by effectively using spatial and temporal features of mesh data, allowing for the construction of a highly accurate flow rate prediction model.
Smart Images

Figure 2024167605000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to a prediction device and a prediction method. [Background technology]
[0002] Conventionally, river flow rates have been predicted based on meteorological information such as rainfall amounts, for example, to prepare for disasters caused by typhoons, heavy rain, etc. Prediction models generated by machine learning are sometimes used to predict river flow rates.
[0003] In flow rate prediction using a prediction model, a prediction model is generated from past rainfall and weather forecast information, for example, by machine learning using a neural network. That is, a prediction model is constructed to predict river flow rate from data such as rainfall at points within the river's catchment area, and future river flow rate is predicted by inputting data such as actual rainfall into this prediction model. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2020-111892 A [Patent Document 2] JP 2022-156802 A [Patent Document 3] JP 2007-205001 A Summary of the Invention [Problem to be solved by the invention]
[0005] However, in flow rate forecasts using a prediction model, the data used to construct the prediction model and make predictions using the prediction model is limited, and there is a problem in that there is a certain limit to improving the prediction accuracy. Specifically, although the flow rate of a river is affected by the amount of rainfall in the entire catchment area, only data such as the amount of rainfall at specific points in the catchment area is used to construct the prediction model and make predictions, so the rainfall conditions in the entire catchment area are not reflected in the prediction results. As a result, it may be difficult to sufficiently improve the accuracy of the flow rate forecast.
[0006] In general, rainfall amount and the like for an entire catchment area can be obtained from mesh-format data. Therefore, if mesh data can be used for machine learning, it is possible to improve the accuracy of prediction models and prediction results. However, mesh data is usually very large-scale data, and it is not easy to introduce mesh data into the construction of prediction models or predictions using the prediction models.
[0007] The present disclosure has been made in consideration of the above, and aims to provide a prediction device and a prediction method that can improve prediction accuracy using a prediction model constructed by machine learning. [Means for solving the problem]
[0008] According to one aspect of the present disclosure, a prediction device has an acquisition unit that acquires data including mesh data, a learning unit that inputs the data acquired by the acquisition unit into a prediction model to learn model parameters, and a prediction unit that performs predictions based on the data acquired by the acquisition unit using the prediction model in which the model parameters learned by the learning unit are set, and the learning unit learns model parameters of a prediction model that has a spatial feature extractor that extracts spatial features of the input mesh data and a temporal feature extractor that extracts temporal features of the data input from the spatial feature extractor.
[0009] Moreover, according to one aspect of the present disclosure, a prediction method includes steps of acquiring data including mesh data, inputting the acquired data into a prediction model to learn model parameters, and performing prediction based on the acquired data using the prediction model in which the learned model parameters are set, wherein the learning step learns model parameters of a prediction model having a spatial feature extractor that extracts spatial features of the input mesh data and a temporal feature extractor that extracts temporal features of the data input from the spatial feature extractor. [Brief description of the drawings]
[0010] [Figure 1] FIG. 1 is a block diagram showing a configuration of a flow rate prediction device according to an embodiment. [Diagram 2] FIG. 2 is a block diagram showing the configuration of the learning unit. [Diagram 3] FIG. 3 is a block diagram showing the configuration of the prediction unit. [Figure 4] FIG. 4 is a diagram showing the configuration of a flow rate prediction model. [Diagram 5] FIG. 5 is a flow diagram illustrating a flow rate prediction method according to an embodiment. [Figure 6] FIG. 6 is a flow diagram showing the operation of the flow rate prediction model during learning and prediction. [Figure 7] FIG. 7 is a diagram illustrating an example of a data format of the multidimensional array data. [Figure 8] FIG. 8 is a diagram illustrating the first mini-batch processing method. [Figure 9] FIG. 9 illustrates pseudocode corresponding to the first mini-batch processing method. [Figure 10] FIG. 10 is a diagram illustrating the second mini-batch processing method. [Figure 11] FIG. 11 illustrates pseudocode corresponding to the second mini-batch processing method. [Figure 12] FIG. 12 is a diagram illustrating the third mini-batch processing method. [Figure 13]FIG. 13 illustrates pseudocode corresponding to the third mini-batch processing method. [Figure 14] FIG. 14 is a block diagram showing a specific example of a hardware configuration of a flow rate prediction device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0011] Hereinafter, an embodiment according to the present disclosure will be described with reference to the accompanying drawings. The embodiment described below is merely an example, and should not be construed as being limited by this description.
[0012] Fig. 1 is a block diagram showing the configuration of a flow rate prediction device 100 according to an embodiment. This flow rate prediction device 100 predicts the flow rate in a river based on various meteorological data. The flow rate prediction device 100 shown in Fig. 1 includes a data acquisition unit 110, a data storage unit 120, a setting unit 130, a data preprocessing unit 140, a processed data storage unit 150, a learning unit 160, a parameter storage unit 170, and a prediction unit 180.
[0013] The data acquisition unit 110 acquires various data used for learning the flow rate prediction model and predicting the flow rate. Specifically, the data acquisition unit 110 acquires point record data, point flow rate data, mesh record data, and numerical forecast data.
[0014] The point record data is meteorological and hydrological data obtained using a measuring instrument at a certain measurement point, and is, for example, two-dimensional array data of time and item. The point flow rate data is data in which the flow rate of the river to be predicted is recorded, and is, for example, two-dimensional array data of time and variable name. The mesh record data is mesh-divided data such as geographical altitude and meteorological information observed by weather radar, and is, for example, four-dimensional array data of time, latitude, longitude, and variable name. The numerical forecast data is numerical simulation data calculated using a computer by dividing the atmosphere into a grid in order to forecast future weather, and is, for example, six-dimensional array data of time, predicted time, latitude, longitude, altitude, and variable name. The mesh record data and the numerical forecast data are examples of mesh data having attribute data for each mesh, and are large-scale data.
[0015] The data acquiring unit 110 may acquire these data from other servers via a network, or may acquire the data by accepting input of data measured by a measuring instrument.
[0016] The data storage unit 120 stores various data acquired by the data acquisition unit 110 .
[0017] The setting unit 130 executes various settings related to preprocessing of data, learning of a flow prediction model, and prediction of flow using the flow prediction model. Specifically, when preprocessing is executed to extract data in a range used for learning and prediction, the setting unit 130 notifies the data preprocessing unit 140 of hyperparameters that specify this range. The setting unit 130 also notifies the learning unit 160 of hyperparameters that specify, for example, a mini-batch size and the number of learning epochs when learning the flow prediction model. Furthermore, the setting unit 130 notifies the prediction unit 180 of hyperparameters that specify, for example, a maximum prediction destination time and a mini-batch size when predicting using the flow prediction model.
[0018] The data preprocessing unit 140 performs preprocessing on the data stored by the data storage unit 120 according to the hyperparameters. Specifically, the data preprocessing unit 140 performs trimming to cut out a range related to the prediction target from the mesh record data and the numerical forecast data. That is, the data preprocessing unit 140 cuts out data of a mesh corresponding to the catchment area of the river of the prediction target from the mesh record data and the numerical forecast data. The data preprocessing unit 140 also trims the mesh record data and the numerical forecast data in the time direction by a time width according to the learning period. Furthermore, the data preprocessing unit 140 may delete information on altitude from the numerical forecast data when, for example, weather information of the upper atmosphere is not necessary for prediction. The data preprocessing unit 140 may also normalize the necessary data obtained by trimming.
[0019] The processed data storage unit 150 stores various data preprocessed by the data preprocessing unit 140. In addition, the processed data storage unit 150 generates mini-batches, which are groups of data used for learning a flow prediction model and predicting a flow rate, in response to requests from the learning unit 160 and the prediction unit 180, and outputs the mini-batches to the learning unit 160 and the prediction unit 180.
[0020] The learning unit 160 learns the flow prediction model using data output from the processed data storage unit 150 in mini-batch units. The window size, maximum prediction destination time, numerical value indicating the structure of the flow prediction model, mini-batch size, number of learning epochs, learning algorithm to be used, and the like related to learning in the learning unit 160 are designated as hyperparameters by the setting unit 130. The learning unit 160 learns the flow prediction model which is a regression model using a neural network. At this time, the learning unit 160 learns the flow prediction model which converts the mesh record data and numerical forecast data, which are multidimensional array data, into two-dimensional array data by extracting spatial features, and extracts temporal features from the converted data to predict the flow rate. The flow prediction model learned by the learning unit 160 will be described in detail later.
[0021] The parameter storage unit 170 stores the model parameters of the flow rate prediction model learned by the learning unit 160.
[0022] The prediction unit 180 constructs a flow rate prediction model using the model parameters stored by the parameter storage unit 170, and inputs data output from the processed data storage unit 150 in mini-batch units to the flow rate prediction model to predict the flow rate. The maximum prediction destination time and mini-batch size for prediction in the prediction unit 180 are specified as hyperparameters by the setting unit 130. The flow rate prediction model used by the prediction unit 180 to predict the flow rate will be described in detail later.
[0023] 2 is a block diagram showing the configuration of the learning unit 160. The learning unit 160 shown in FIG.
[0024] The learning management unit 161 plans and manages the learning progress of the flow rate prediction model according to the hyperparameters notified by the setting unit 130. That is, the learning management unit 161 outputs information about the mini-batches used in learning to the mini-batch processing unit 162. In addition, when learning using mini-batches is executed by the mini-batch learning unit 163, the learning management unit 161 acquires model parameters of the flow rate prediction model obtained by learning, and stores them in the parameter storage unit 170.
[0025] The mini-batch processing unit 162 determines the data format of the mini-batch to be used for learning, and acquires a data entity corresponding to the mini-batch of the determined data format from the processed data storage unit 150. Then, the mini-batch processing unit 162 converts the data entity acquired from the processed data storage unit 150 into mini-batch data in a form suitable for mini-batch learning, and outputs the converted data to the mini-batch learning unit 163.
[0026] The mini-batch learning unit 163 executes mini-batch learning using the mini-batch data input from the mini-batch processing unit 162, and optimizes the model parameters of the flow prediction model. That is, the mini-batch learning unit 163 inputs the point record data and the mini-batch data to the flow prediction model to calculate a predicted flow rate, and optimizes the model parameters of the flow prediction model using an error backpropagation method that backpropagates the error between the calculated predicted flow rate and the point flow rate data. Then, the mini-batch learning unit 163 notifies the learning management unit 161 of the optimized model parameters.
[0027] 3 is a block diagram showing the configuration of the prediction unit 180. The prediction unit 180 shown in FIG.
[0028] The prediction management unit 181 plans and manages a prediction progress plan using the flow rate prediction model according to the hyperparameters notified from the setting unit 130. That is, the prediction management unit 181 outputs information about the mini-batches used for prediction to the mini-batch processing unit 182. In addition, when the mini-batch prediction unit 183 executes a prediction using the flow rate prediction model, the prediction management unit 181 outputs a predicted flow rate, which is the prediction result.
[0029] The mini-batch processing unit 182 determines the data format of the mini-batch to be used for prediction, and acquires a data entity corresponding to the mini-batch of the determined data format from the processed data storage unit 150. Then, the mini-batch processing unit 182 converts the data entity acquired from the processed data storage unit 150 into mini-batch data in a form that can be input to a flow rate prediction model, and outputs the data to the mini-batch prediction unit 183.
[0030] The mini-batch prediction unit 183 acquires model parameters of the flow prediction model from the parameter storage unit 170, and inputs the mini-batch data input from the mini-batch processing unit 182 to the flow prediction model to execute prediction. That is, the mini-batch prediction unit 183 inputs the point record data and the mini-batch data to the flow prediction model in which the model parameters read from the parameter storage unit 170 are set, and calculates a predicted flow rate. Then, the mini-batch prediction unit 183 notifies the prediction management unit 181 of the calculated predicted flow rate.
[0031] Next, a description will be given of a flow rate prediction model that is learned by the mini-batch learning unit 163 and used for prediction in the mini-batch prediction unit 183. Fig. 4 is a diagram showing a configuration of a flow rate prediction model according to this embodiment. The flow rate prediction model has spatial feature extractors 210 and 230, an encoder 220, and a decoder 240.
[0032] The spatial feature extractor 210 receives as input data the point record data and point flow rate data from L time ago (L is an integer equal to or greater than 1) to the present (or the learning reference time) and the mesh record data from L time ago to the present (or the learning reference time) contained in the mini-batch data, and extracts spatial features of the input data. Note that the length of time L from L time ago to the present (or the learning reference time) is the time width of the data used for learning, and is also called the window size. In addition, if the mini-batch data contains numerical forecast data, the numerical forecast data from L time ago to the present (or the learning reference time) may be used as input data to the spatial feature extractor 210.
[0033] The spatial feature extractor 210 extracts features of the mesh record data in a spatial feature extraction layer, and converts the multidimensional array of mesh record data into a two-dimensional array of time and item data. Such a spatial feature extraction layer can be realized by, for example, a Convolutional Neural Network (CNN) mechanism or an Attention mechanism.
[0034] The spatial feature extractor 210 combines the two-dimensional array of point record data and point flow rate data with the mesh record data converted to a two-dimensional array in a combination layer. As a result, the spatial feature extractor 210 extracts spatial features from the input data to generate two-dimensional array data, and outputs this two-dimensional array data to the encoder 220.
[0035] The encoder 220 is a temporal feature extractor that receives two-dimensional array data output from the spatial feature extractor 210 as input data and extracts temporal features of the input data. That is, the encoder 220 extracts features in the temporal direction from data for each time from L time points before to the present (or the learning reference time) in a temporal feature extraction layer. Such a temporal feature extraction layer can be realized by, for example, a recurrent neural network (RNN) mechanism or an attention mechanism. The encoder 220 outputs the extracted features to the decoder 240.
[0036] The spatial feature extractor 230 receives as input data the numerical forecast data from one time ahead of the present (or the learning reference time) to M times ahead (M is an integer of 2 or more) included in the mini-batch data and the predicted flow rate of the numerical forecast data one time before for each time, and extracts spatial features of the input data. During learning, point flow rate data can be used as the predicted flow rate at each time. During prediction, for the time one time ahead of the present, since there is no predicted flow rate one time before (i.e., the present), 0 is input as the predicted flow rate.
[0037] The spatial feature extractor 230 extracts features of the numerical forecast data in a spatial feature extraction layer, and converts the multidimensional array of the numerical forecast data into a two-dimensional array of data of time and item. Such a spatial feature extraction layer can be realized by, for example, a convolutional neural network (CNN) mechanism or an attention mechanism.
[0038] The spatial feature extractor 230 combines the two-dimensional array of predicted flow rates with the numerical forecast data converted into a two-dimensional array in a combination layer. As a result, the spatial feature extractor 230 extracts spatial features from the input data to generate two-dimensional array data, and outputs this two-dimensional array data to the decoder 240.
[0039] The decoder 240 is a time feature extractor that receives the feature values output from the encoder 220 and the two-dimensional array data output from the spatial feature extractor 230 as input data and extracts the time feature values of the input data. That is, in a time feature extraction layer, the decoder 240 extracts time-direction feature values from the feature values output from the encoder 220 and data for each time from the present (or learning reference time) to M time points ahead. Such a time feature extraction layer can be realized by, for example, a recurrent neural network (RNN) mechanism or an attention mechanism. The decoder 240 outputs a predicted flow rate for each time by combining the feature values in the time direction in a fully connected layer.
[0040] By performing learning and prediction using such a flow rate prediction model, it is possible to introduce large-scale mesh data such as mesh record data and numerical forecast data, thereby improving the prediction accuracy using a prediction model built by machine learning.
[0041] Next, a flow rate prediction method by the flow rate prediction device 100 configured as above will be described with reference to a flow chart shown in FIG.
[0042] First, the data acquisition unit 110 acquires various data including point record data, point flow rate data, mesh record data, and numerical forecast data (step S101). The acquired data is stored in the data storage unit 120. Then, the data preprocessing unit 140 executes preprocessing on the data stored in the data storage unit 120 (step S102). The contents of the preprocessing are set by hyperparameters notified by the setting unit 130. The preprocessed data is stored in the processed data storage unit 150.
[0043] Then, the learning unit 160 acquires the mini-batch data from the processed data storage unit 150, and executes learning of the flow rate prediction model using the mini-batch data (step S103). Specifically, the mini-batch processing unit 162 acquires the mini-batch data cut out from the mesh record data and the numerical forecast data, and the mini-batch learning unit 163 optimizes the model parameters of the flow rate prediction model by machine learning using the mini-batch data. The optimized model parameters are stored in the parameter storage unit 170.
[0044] Then, the prediction unit 180 acquires the mini-batch data from the processed data storage unit 150, and inputs the mini-batch data to the flow rate prediction model to execute prediction (step S104). Specifically, the mini-batch processing unit 182 acquires the mini-batch data extracted from the mesh record data and the numerical forecast data, and the mini-batch prediction unit 183 inputs the mini-batch data to the flow rate prediction model in which the optimized model parameters are set, and calculates the predicted flow rate.
[0045] In this way, learning and prediction are performed using mini-batch data based on mesh data such as mesh record data and numerical forecast data, so that a highly accurate flow prediction model can be constructed and prediction accuracy can be improved.
[0046] Next, the operations during learning and prediction using mini-batch data will be described with reference to the flow diagram shown in FIG.
[0047] During learning and prediction, the mini-batch processing unit 162 or the mini-batch processing unit 182 performs mini-batch processing to process mesh data into mini-batch data (step S201). Specifically, an index of data to be used as mini-batch data is determined from the mesh record data and the numerical forecast data according to the window size and mini-batch size specified by the hyperparameter. Then, data of the determined index is obtained from the processed data storage unit 150 and converted into mini-batch data.
[0048] The mini-batch data is input to a flow prediction model in the mini-batch learning unit 163 or the mini-batch prediction unit 183. The mini-batch data generated from the mesh record data is then input to the spatial feature extractor 210, which extracts spatial features of the mesh record data (step S202). The extracted feature data is output to the encoder 220, which extracts temporal features of the mesh record data (step S203). The extracted features are output to the decoder 240.
[0049] Meanwhile, the mini-batch data generated from the numerical forecast data is input to the spatial feature extractor 230, which extracts spatial features of the numerical forecast data (step S204). The extracted feature data is output to the decoder 240, which extracts temporal features of the features output from the encoder 220 and the numerical forecast data (step S205). Then, the predicted flow rate for each hour is output from the extracted features (step S206).
[0050] During learning, the model parameters of the flow rate prediction model are optimized by the backpropagation method, which backpropagates the error between the output predicted flow rate and the point flow rate data. During prediction, the output predicted flow rate is output as the prediction result.
[0051] In this way, during learning and prediction, spatial features of mesh data such as mesh record data and numerical forecast data are extracted, and then a flow prediction model is used in which temporal features are extracted. This makes it possible to introduce large-scale mesh data into the flow prediction model. As a result, a highly accurate flow prediction model can be constructed, and prediction accuracy can be improved.
[0052] Next, the generation of mini-batch data to be input to the flow rate prediction model will be described with a specific example. The mini-batch data is generated by the mini-batch processors 162 and 182 acquiring data from the processed data storage unit 150.
[0053] The processed data storage unit 150 stores mesh data such as mesh record data and numerical forecast data. The data format of the mesh data stored in the processed data storage unit 150 can be expressed, for example, as shown in Fig. 7. That is, the processed data storage unit 150 stores multidimensional data represented by indexes such as "D01:00" for each date and time. The mini-batch data is obtained by extracting mini-batches of consecutive time widths of the window size from such mesh data, and includes the number of mini-batches corresponding to the mini-batch size.
[0054] For example, consider a case where the window size is equivalent to three time periods, and mini-batch data is generated consisting of mini-batches corresponding to three time periods, "2000 / 01 / 01 03:00", "2000 / 01 / 01 07:00", and "2000 / 01 / 01 04:00", selected by a predetermined sampling. When generating the mini-batch data, the mini-batch processing units 162 and 182 determine an index of data to be used as the mini-batch data, and obtain a data entity of the determined index from the processed data storage unit 150. Then, the mini-batch processing units 162 and 182 generate the mini-batch data from the obtained data entity.
[0055] In the first method of generating mini-batch data, data is generated by shifting mesh data by the window size, and a mini-batch selected by sampling is extracted from the shifted data. Specifically, as shown in the left diagram of Fig. 8, for example, mesh data stored in the processed data storage unit 150 is first used as data at a reference time, and data at the reference time is shifted to generate data one time before and two time before. This generates data with a continuous time width of three time periods, which is the window size.
[0056] Then, for example, as shown in the right diagram of FIG. 8, mini-batches corresponding to three times selected by predetermined sampling are extracted from the shifted data, thereby generating mini-batch data.
[0057] Pseudocode corresponding to the first method is shown in FIG. 9. In this pseudocode, code 311 indicates a process of shifting data at a reference time by the window size L times. Code 312 indicates a process of combining the shifted data. And code 313 indicates a process of extracting a number of mini-batches corresponding to the mini-batch size from the shifted data. According to such a first method, mini-batch data can be generated by the simple process of shifting data and extracting mini-batches, so that mini-batch data can be generated with a small amount of calculation processing.
[0058] In the second method of generating mini-batch data, data of a window size width is extracted from mesh data, the extracted data is transformed to swap rows and columns, and the transformed data is concatenated in the row direction. Specifically, as shown in the left diagram of FIG. 10, data of a window size width based on each of three times selected by a predetermined sampling is extracted. That is, in this example, data of the reference time "2000 / 01 / 01 03:00" and data of one hour and two hours before this reference time are extracted, and data of the reference time, one hour and two hours before the reference time are extracted for other reference times.
[0059] Then, for example, as shown in the right diagram of Fig. 10, shape transformation is performed to convert the data arranged in the column direction extracted for each reference time into data arranged in the row direction. Furthermore, the shape-transformed data is concatenated in the row direction to generate mini-batch data.
[0060] Pseudocode corresponding to the second method is shown in FIG. 11. In this pseudocode, code 321 indicates a process of extracting data of a window size L width for each of three reference times corresponding to the mini-batch size B. Code 322 indicates a process of shape-transforming the data extracted for each reference time. Code 323 indicates a process of concatenating the shape-transformed data in the row direction. According to such a second method, data obtained by shifting the mesh data generated in the first method is not required, and it is possible to eliminate the need for a memory for storing large amounts of data generated as intermediate products.
[0061] In the third method of generating mini-batch data, data of the mini-batch size is extracted from the mesh data, the date and time corresponding to the extracted data is replaced with the date and time of the reference time, and data corresponding to the same date and time is concatenated in the column direction. Specifically, as shown in the left diagram of FIG. 12, data of the mini-batch size corresponding to three times selected by predetermined sampling is extracted. That is, here, for example, data of "2000 / 01 / 01 01:00", "2000 / 01 / 01 05:00", and "2000 / 01 / 01 02:00", which are two hours before the three reference times, are extracted, and three pieces of data equal to the mini-batch size are extracted for one hour before the reference time and for the reference time, respectively.
[0062] Then, for example, as shown in the right diagram of Fig. 12, the date and time of each data extracted for each time period is replaced with the date and time of the reference time. Furthermore, the data is concatenated in the column direction using the replaced date and time as a key to generate mini-batch data.
[0063] Pseudocode corresponding to the third method is shown in FIG. 13. In this pseudocode, code 331 indicates a process of extracting data of mini-batch size B for each of three times included in the window size L. Code 332 indicates a process of replacing the date and time of the data extracted for each time with the date and time of the reference time. Code 333 indicates a process of concatenating the data whose date and time have been replaced in the column direction. According to such a third method, since data for each time included in the window size is concatenated, the number of times data is concatenated does not increase even if the mini-batch size becomes large. Therefore, compared to the second method, the amount of calculation processing for generating mini-batch data can be reduced, and mini-batch data can be generated at high speed.
[0064] As described above, according to this embodiment, a spatial feature extractor extracts spatial features of mesh data, and the resulting two-dimensional array data is used as input data for the encoder and decoder to learn a flow prediction model that extracts temporal features, and the flow rate is predicted using this flow prediction model. Therefore, large-scale mesh data with a large amount of information, such as mesh record data and numerical forecast data, can be used for machine learning. As a result, a highly accurate prediction model can be constructed by machine learning, and prediction accuracy can be improved.
[0065] The flow rate prediction device 100 according to the embodiment can be configured using a processor and a memory. Fig. 14 is a block diagram showing an example of a hardware configuration of the flow rate prediction device 100 according to the embodiment. As shown in Fig. 14, the flow rate prediction device 100 includes a processor 101, a memory 102, an input / output unit 103, and a storage 104.
[0066] The processor 101 has, for example, a central processing unit (CPU), a field programmable gate array (FPGA), or a digital signal processor (DSP), and controls the entire flow rate prediction device 100 and executes various types of arithmetic processing.
[0067] The memory 102 includes, for example, a random access memory (RAM) or a read only memory (ROM), and stores information used in the arithmetic processing executed by the processor 101.
[0068] The input / output unit 103 is an interface through which the user inputs information and outputs information to the user, and may include, for example, a keyboard, a display, a touch panel, a microphone, or a speaker.
[0069] The storage 104 includes, for example, a hard disk drive (HDD) or a solid state drive (SSD), and stores various types of data.
[0070] The processing by the flow rate prediction device 100 described in the above embodiment can also be written as a computer-executable program. In this case, the program can be stored in a computer-readable and non-transitory recording medium and installed in the computer. Examples of such recording media include portable recording media such as CD-ROMs, DVD disks, and USB memories, and semiconductor memories such as flash memories.
[0071] In addition, in the above embodiment, a flow prediction model for predicting the flow rate of a river was used as an example, but the learning and prediction processes related to the present disclosure can be applied to various prediction models other than the flow prediction model. [Explanation of symbols]
[0072] 101 Processor 102 Memory 103 Input / output section 104 Storage 110 Data Acquisition Section 120 Data storage unit 130 Setting section 140 Data Preprocessing Unit 150 Processing data storage unit 160 Learning Department 161 Learning Management Department 162, 182 Mini batch processing section 163 Mini-batch learning part 170 Parameter storage section 180 Prediction Department 181 Forecasting Management Department 183 Mini-batch prediction part 210, 230 Spatial feature extractor 220 Encoder 240 Decoder
Claims
1. An acquisition unit that acquires data including mesh data; a learning unit that inputs the data acquired by the acquisition unit into a prediction model to learn model parameters; a prediction unit that executes a prediction based on the data acquired by the acquisition unit, using a prediction model in which the model parameters learned by the learning unit are set; The learning unit is a spatial feature extractor for extracting spatial features of input mesh data; and a time feature extractor that extracts a time feature amount of the data input from the spatial feature extractor. Prediction device.
2. The acquisition unit is Obtaining meteorological mesh data, The learning unit is inputting the mesh data into a flow rate prediction model for predicting the flow rate of a river to learn model parameters; The prediction unit is Using the flow prediction model, perform a prediction of flow based on the mesh data. The prediction device according to claim 1 .
3. The learning unit is a mini-batch processing unit that acquires multi-dimensional data corresponding to a date and time determined by sampling from mesh data that holds multi-dimensional data associated with a date and time, and generates mini-batch data by shape-converting or replacing the acquired date and time and multi-dimensional data; a mini-batch learning unit that inputs mini-batch data generated by the mini-batch processing unit into the prediction model to learn model parameters; The prediction device according to claim 1 , further comprising:
4. The prediction unit is a mini-batch processing unit that acquires multi-dimensional data corresponding to a predetermined date and time from mesh data that holds multi-dimensional data associated with a date and time, and generates mini-batch data by shape-converting or replacing the acquired date and time and multi-dimensional data; a mini-batch prediction unit that inputs the mini-batch data generated by the mini-batch processing unit into the prediction model and executes prediction; The prediction device according to claim 3 , further comprising:
5. Obtain data including mesh data, The acquired data is input into a predictive model to learn the model parameters, Executing a prediction based on acquired data using a prediction model in which the learned model parameters are set; The learning step includes: a spatial feature extractor for extracting spatial features of input mesh data; and a time feature extractor that extracts a time feature amount of the data input from the spatial feature extractor. Forecasting methods.
Citation Information
Patent Citations
Discharge forecasting apparatus
JP2007205001A
Flow rate prediction device and flow rate prediction method
JP2020111892A
Flow prediction program, flow prediction method and information processing device
JP2022156802A
