Flow rate prediction device and flow rate prediction method

Abstract

To provide a flow rate prediction device and a flow rate prediction method capable of predicting a flow rate with high accuracy by extracting important features.SOLUTION: And an ensemble learning processing unit 15 that learns an ensemble model so as to minimize an error between a flow rate prediction value and an actual value based on the important feature quantity data extracted by the feature quantity extraction unit 11, and a flow rate prediction unit 19 that constructs a prediction model using a parameter of the ensemble model learned by the ensemble learning processing unit 15, and predicts a flow rate based on the important feature quantity data extracted by the feature quantity extraction unit 11.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a flow rate prediction device and a flow rate prediction method. [Background technology]

[0002] Mid- to long-term changes in river flow affect dam power generation and water release plans. Sensors are installed in several locations upstream of dams to acquire real-time data on flow rate, temperature, and rainfall. These data are analyzed in conjunction with weather forecast data for the relevant region obtained from the Japan Meteorological Agency to predict changes in inflow 24 to 48 hours in the future. To automate this flow rate forecasting, time-series forecasting using machine learning has been proposed in recent years. Specifically, attempts have been made to generate a predictive model using machine learning and use it to forecast flow rate. For example, Patent Document 1 discloses a method for predicting inflow using machine learning based on past rainfall and weather forecast information. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-111892 [Non-patent literature]

[0004] [Non-Patent Document 1] G Ke, Q Meng, T Finley, T Wang, W Chen, W Ma, Q Ye, TY Liu, “LightGBM: A Highly Efficient Gradient Boosting Decision Tree”, Advances in neural information processing systems, 2017 Summary of the Invention [Problem to be solved by the invention]

[0005] In general, time series forecasting using machine learning can be affected by noise and uncorrelated data during the learning process. Furthermore, because regional forecast mesh data related to flow rate prediction, such as rainfall, temperature, pressure, and humidity, have multiple dimensions, inputting them as is increases the computational costs of the model. Therefore, efficient flow rate forecasting using machine learning requires feature engineering, which removes noise, missing data, and irrelevant data from data from multiple locations and types, and extracts useful features from high-dimensional data.

[0006] The flow rate prediction device and method disclosed in Patent Document 1 are a method for predicting flow rate by inputting various data into a deep learning model. With this method, there is a risk that noise or irrelevant data may be mixed into the input data, and there is a challenge in how to prevent the mixing of this irrelevant data and improve the prediction accuracy of the AI ​​model used.

[0007] The present invention has been made in consideration of the above-mentioned circumstances, and the problem that the present invention aims to solve is to provide a flow rate prediction device and a flow rate prediction method that enable highly accurate flow rate prediction by extracting important features. [Means for solving the problem]

[0008] In order to solve the above problems, the present invention employs the following means. In other words, the flow rate prediction device of the present invention is a flow rate prediction device that predicts the flow rate of a specific river, and comprises: a feature importance analysis unit that detects the importance of features from observation data and meteorological mesh data; a feature extraction unit that extracts important features from the observation data and meteorological mesh data using parameters related to the detection of the importance of the features in the feature importance analysis unit; an ensemble learning processing unit that learns an ensemble model so as to minimize the error between the flow rate prediction value and the actual value based on the important feature data extracted by the feature extraction unit; and a flow rate prediction unit that constructs a prediction model using the parameters of the ensemble model learned in the ensemble learning processing unit, and predicts the flow rate based on the important feature data extracted by the feature extraction unit.

[0009] The flow rate prediction method of the present invention is a flow rate prediction method for predicting the flow rate of a specific river, and includes detecting the importance of features from observation data and meteorological mesh data, extracting important features from the observation data and meteorological mesh data using parameters related to detecting the importance of the features, training an ensemble model so as to minimize the error between the flow rate prediction value and the actual value based on the important feature data extracted by extracting the important features, constructing a prediction model using the parameters of the trained ensemble model, and predicting the flow rate based on the important feature data extracted by extracting the important features. [Effects of the Invention]

[0010] According to the present invention, it is possible to provide a flow rate prediction device and a flow rate prediction method that are capable of highly accurate flow rate prediction by extracting important features. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a block diagram showing a configuration of a flow rate prediction device according to an embodiment of the present invention. [Figure 2] FIG. 10 is a visualization diagram of a decision tree in a feature importance analysis unit according to an embodiment of the present invention. [Figure 3]FIG. 2 is a configuration diagram of an important feature extraction unit in a feature importance analysis unit according to an embodiment of the present invention. [Figure 4] FIG. 2 is a configuration diagram of an important feature extraction unit in a feature importance analysis unit according to an embodiment of the present invention. [Figure 5] This is a diagram showing the configuration of the CNN in Figures 3 and 4. [Figure 6] FIG. 2 is a configuration diagram of ensemble learning in an ensemble learning processing unit according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0012] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A flow rate prediction device according to the present invention is a flow rate prediction device for predicting the flow rate of a specific river.

[0013] (Embodiment) Fig. 1 is a block diagram showing the configuration of a flow rate prediction device 1 according to an embodiment of the present invention. As shown in Fig. 1, the flow rate prediction device 1 includes a data storage unit 3, a data processing unit 5, a feature importance analysis unit 7, an analysis parameter storage unit 9, a feature extraction unit 11, an important feature storage unit 13, an ensemble learning processing unit 15, a model parameter storage unit 17, and a flow rate prediction unit 19.

[0014] The data storage unit 3 stores observation data and meteorological mesh data. The data processing unit 5 processes the data stored in the data storage unit 3 by filling in missing data, converting the format, normalizing each element, and so on. The feature importance analysis unit 7 detects the importance of features from the data processed by the data processing unit 5. The analysis parameter storage unit 9 stores parameters related to feature importance detection in the feature importance analysis unit 7. The feature extraction unit 11 extracts important features from the data processed by the data processing unit 5 using the parameters related to feature importance detection stored in the analysis parameter storage unit 9. The important feature storage unit 13 stores the important feature data extracted by the feature extraction unit 11. The ensemble learning processing unit 15 learns an ensemble model based on the important feature data stored in the important feature storage unit 13 to minimize the error between the predicted flow rate value and the actual value. The model parameter storage unit 17 stores the parameters of the ensemble model learned by the ensemble learning processing unit 15. The flow rate prediction unit 19 constructs a prediction model using the parameters of the ensemble model stored in the model parameter storage unit 17, and predicts the flow rate based on the important feature data extracted by the feature extraction unit 11.

[0015] The flow rate prediction device 1 is an information processing device such as a personal computer or a server configured on the cloud. Among the components of the flow rate prediction device 1, the data processing unit 5, the feature importance analysis unit 7, the feature extraction unit 11, the ensemble learning processing unit 15, and the flow rate prediction unit 19 may be software or programs executed by a processor such as a CPU, a GPU (Graphics Processing Unit), or a TPU (Tensor Processing Unit) within the information processing device. Furthermore, the data storage unit 3, the analysis parameter storage unit 9, the important feature storage unit 13, and the model parameter storage unit 17 may be realized by storage devices such as semiconductor memories or magnetic disks provided inside or outside the information processing device.

[0016] Next, the operation of the flow rate prediction device configured as above will be described. As shown in Fig. 1, observation data and meteorological mesh data are stored in the data storage unit 3. Here, the observation data consists of past hydrological data such as river water levels and flow rates upstream of dams, which are collected periodically from local hydrological observation stations and water management organizations, and past sensor data such as temperature, humidity, rainfall, and dam inflow, which are collected periodically from meteorological and flow rate sensors installed around dams and rivers.

[0017] Meteorological mesh data is provided by meteorological organizations such as the Japan Meteorological Agency. This is meteorological data on a spatial scale of several kilometers to several tens of kilometers, generated by a weather forecast model, with values ​​of meteorological elements (rainfall, temperature, air pressure, wind speed, humidity, etc.) stored in each mesh cell. It is also multidimensional weather forecast data with a time scale ranging from several hours to several days, depending on the weather model. The data storage unit 3 stores the above observation data and meteorological mesh data, linked to date data such as time, season, and year, month, and day associated with each data.

[0018] The data processing unit 5 reads the data stored in the data storage unit 3, standardizes each data, fills in missing data, and converts it into a format that can be input to an AI model. Observation data without a spatial axis is converted into time-ordered numeric table data, and meteorological mesh data with both space and time is converted into time-ordered matrix format data. Date data such as time, season, and year, month, and day is converted into embedded representations, and then converted into numeric table data together with the observation data.

[0019] The feature importance analysis unit 7 uses an algorithm to analyze the importance of features of the data converted by the data processing unit 5. For observation data that does not have a spatial axis, a gradient boosting regression tree (GBRT) algorithm in machine learning is used to quantify the importance of features. In learning the GBRT algorithm, parameters are adjusted by minimizing the error between the output value and the actual flow rate value.

[0020] In adjusting this parameter, the gain is calculated according to the formula 1 shown below when dividing at each branch node of the decision tree, and the extent to which the error in flow rate prediction is reduced is evaluated. L and G R , H L and H R are the gradient and Hessian of the objective function at the divided left and right child nodes. λ and γ are regularization parameters. Figure 2 shows a visualized diagram of the decision tree of LightGBM, a representative algorithm of the trained GBRT system. For this analysis, for example, the method disclosed in Non-Patent Document 1 is used.

[0021]

number

[0022] In Figure 2, it can be seen that the larger the gain, the more the feature in question is concentrated at the top of the decision tree and is frequently used in branching nodes (current_inflow, shown by the dashed ellipse in Figure 2). This allows the trained LightGBM to quantitatively evaluate the extent to which the feature contributes to prediction, sorting them in descending order, and extracting the top n features as important features, where n is a variable integer.

[0023] On the other hand, for meteorological mesh data that has both spatial and temporal axes, an algorithm called a convolutional neural network (CNN) is used in deep learning to extract important features, as shown in Figures 3 and 4. Figure 5 shows the configuration of the above CNN model. The model is configured to extract features from the input mesh data using multiple convolutional layers and generate a feature map. A pooling layer is added between the convolutional layers to reduce the size of the feature map. The feature map is then transformed into a one-dimensional feature vector using a flattening layer, and the size of the vector is adjusted using a fully connected layer. Finally, an output layer is constructed based on whether all forecast times are output simultaneously or individually. However, the above CNN model requires parameter adjustment by using actual flow rate values ​​as labels and minimizing the error with the output value.

[0024] The analysis parameter storage unit 9 stores the trained LightGBM and CNN parameters and parameters related to important feature extraction from the feature importance analysis unit 7. The feature extraction unit 11 performs processing to extract important features from the data processed by the data processing unit 5 using the parameters stored in the analysis parameter storage unit 9. The important feature storage unit 13 stores the important feature data extracted by the feature extraction unit 11 as table data for model training.

[0025] The ensemble learning processor 15 adjusts the parameters of the ensemble model by minimizing the difference between the predicted output value and the actual flow rate value using the learning data stored in the important feature memory 13. Figure 6 shows the configuration of ensemble learning. In this embodiment, ensemble learning is based on a starking structure. The base model uses an existing AI model, and the metamodel is constructed using a neural network model customized based on the prediction period and the length of the input vector. The base model combines LightGBM, a GBRT-based model with the highest accuracy in time series prediction tasks; LSTM, an RNN-based model that can effectively learn long-term dependencies; and TSMixer, a state-of-the-art MLP-based time series prediction model developed based on MLP-Mixer. These models complement each other's weaknesses and capture a wider variety of patterns, improving prediction performance. The metamodel uses a neural network to combine the prediction results of different models and output the final flow rate prediction value.

[0026] In this embodiment, the ensemble model has three stages: <Stage 1> Three machine learning models, LightGBM, LSTM, and TSMixer, are used to predict traffic flow at all prediction time points. However, each model uses different training data periods. The models trained in Stage 1 are inferred using the same test set, and the prediction results are stored in the Stage 1 prediction result dataset.

[0027] <Stage 2> We continue to use the three models, LightGBM, LSTM, and TSMixer, and add the prediction result dataset from Stage 1 as new feature data. At each different prediction time point, we randomly select one or more models to use as the prediction model. Then, the prediction results for the test set are stored in the prediction result dataset from Stage 2.

[0028] <Stage 3> The meta-model is trained using the forecast result datasets from Stage 1 and Stage 2. The parameters of this model are adjusted to minimize the error between the model output and the actual flow rate values. However, the forecast period range can be changed depending on the needs.

[0029] The model parameter storage unit 17 stores the parameters of the ensemble model that has completed learning in the ensemble learning processing unit 15. The flow rate prediction unit 19 reads the parameters of the learned ensemble model from the model parameter storage unit 17, reconstructs a prediction model, and predicts the flow rate based on the important feature data extracted by the feature extraction unit 11.

[0030] As described above, in this embodiment, the feature importance analysis unit 7 analyzes the importance of features by extracting important features based on a decision tree. Based on this analysis, the feature extraction unit 11 extracts important features from the observation data and meteorological mesh data, and then trains an ensemble model based on these important features, and performs flow rate prediction using a prediction model. Therefore, it is possible to provide a flow rate prediction device and a flow rate prediction method that enable highly accurate flow rate prediction by extracting important features correlated with prediction from high-dimensional data.

[0031] Ensemble learning combines multiple models, enabling more accurate and robust predictions than a single model. However, there is a problem in that the computational cost of multiple models increases exponentially as the number of input features increases. In this embodiment, the combination of important feature extraction processing and ensemble models reduces the number of learning features, so computational costs do not increase significantly even when a hybrid model combining multiple models is used for ensemble learning. Furthermore, by building an ensemble model using representative models in the field of machine learning that are currently expected to have high predictive accuracy as base models, the weaknesses of each model can be compensated for, enabling the capture of more diverse patterns, thereby further improving predictive performance. In other words, by reducing the learning cost of the flow rate prediction AI, it is possible to provide a flow rate prediction device and a flow rate prediction method based on ensemble learning that combines multiple models to achieve more accurate and robust predictions than a single model.

[0032] In addition to this, it is possible to select and discard the configurations given in the above embodiments, or to change them to other configurations as appropriate, without departing from the spirit of the present invention. [Explanation of symbols]

[0033] 1. Flow rate prediction device 3 Data storage unit 5 Data Processing Department 7 Feature Importance Analysis 9 Analysis parameter storage section 11 Feature extraction unit 13 Important feature memory unit 15 Ensemble learning processing unit 17 Model parameter storage section 19 Flow rate prediction section

Claims

1. A flow rate prediction device for predicting the flow rate of a specific river, a feature importance analysis unit that detects the importance of features from the observation data and meteorological mesh data; a feature extraction unit that extracts important feature values ​​from the observation data and the meteorological mesh data using parameters related to feature importance detection in the feature importance analysis unit; an ensemble learning processing unit that learns an ensemble model based on the important feature data extracted by the feature extraction unit so as to minimize an error between a predicted flow rate value and an actual flow rate value; a flow rate prediction unit that constructs a prediction model using parameters of the ensemble model learned in the ensemble learning processing unit, and predicts a flow rate based on important feature data extracted by the feature extraction unit.

2. A flow rate prediction method for predicting the flow rate of a specific river, comprising: Detecting the importance of features from the observation data and meteorological mesh data; extracting important features from the observation data and the meteorological mesh data using parameters related to feature importance detection; learning an ensemble model based on the important feature data extracted by extracting the important feature so as to minimize the error between the predicted flow rate value and the actual value; constructing a prediction model using parameters of the trained ensemble model, and predicting a flow rate based on important feature data extracted by extracting the important features.

Images