Prediction system, prediction method, and program

Abstract

When calculating predicted values ​​using multiple prediction models, the influence of the prediction performance of each individual prediction model is reduced, resulting in more accurate predictions. [Solution] A prediction system that calculates predicted values ​​for data to be predicted using multiple prediction models, comprising: a calculation value calculation unit that calculates calculated values ​​for each prediction model using multiple prediction models that calculate calculated values ​​for predicted values ​​from the values ​​of related data related to the data to be predicted; a weight calculation unit that calculates weights for each prediction model based on the prediction performance of each prediction model; a prediction model extraction unit that extracts one or more prediction models from the multiple prediction models whose most recent prediction performance is nearby as neighboring prediction models; and a prediction value calculation unit that calculates predicted values ​​by inputting the values ​​of related data into the neighboring prediction models and integrating the calculated values ​​calculated for each neighboring prediction model using the weights of each neighboring prediction model.

Description

Technical Field

[0001] The present invention relates to a prediction system, a prediction method, and a program.

Background Art

[0002] Techniques for predicting predicted values of prediction target data that change over time using a prediction model have been conventionally used in various fields (for example, Patent Documents 1 to 3).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0004] Such techniques are also important in the field of electric power. For example, currently, the electric power business is separated into a power generation business and a power transmission business. Along with this, electric power companies are separated into retail traders and power generation companies that conduct free competition, and a power transmission and distribution company that has a neutral position. And the operation of the power system is carried out according to the planned value simultaneous equal quantity system. Specifically, in the planned value simultaneous equal quantity system, retail traders and power generation companies conduct electricity trading according to the supply and demand plans and power generation plans prepared in advance respectively. At this time, if there is an imbalance, which is a deviation between the planned value related to the electricity trading and the actual supply and demand value, the power transmission and distribution company adjusts the supply and demand to operate the power system. Therefore, from the viewpoints of economic rationality and stabilization of the power system, it is important for electric power companies to predict in advance the power demand, the power generation amount of renewable energy utilization type power generation devices, the power market price, etc. using the techniques shown in the above Patent Documents 1 to 3.

[0005] Here, Patent Document 1 describes applying a neural network, which is a machine learning algorithm, to the prediction model to perform prediction using nonlinear approximation.

[0006] Furthermore, Patent Documents 2 and 3 describe methods for calculating more accurate prediction values ​​by integrating the results of predictions made using multiple prediction models. Patent Document 2 describes improving prediction performance through an ensemble effect by aggregating the prediction results of multiple prediction models according to past prediction trends. Specifically, Patent Document 2 describes making predictions using n different prediction methods to calculate n types of prediction values, and using the average of the calculated n types of prediction values ​​as the prediction value for the future.

[0007] Furthermore, Patent Document 3 describes a method for stabilizing prediction performance through outlier removal by selecting one prediction result with the best past error trend from the prediction results of multiple prediction models. Specifically, Patent Document 3 describes constructing a prediction model for calculating predicted energy demand based on actual values ​​of energy demand data and input factor data, and selecting and outputting one of the predicted values ​​calculated using the prediction model according to multiple prediction processing procedures.

[0008] When predictions are made by integrating the prediction results of multiple prediction models, as described in Patent Documents 2 and 3, the prediction performance is often improved compared to when predictions are made using a single prediction model, as described in Patent Document 1. However, as shown in Figures 1 and 2, the appropriate way to integrate the prediction results of each prediction model depends on the prediction performance of those models, so it cannot be said that prediction performance will necessarily improve just because multiple prediction models are used.

[0009] Graph (a) in Figure 1 plots the predicted values ​​of the target data output by the four prediction models A1 to D1 in the past time t0, connected by a dashed line. The horizontal axis represents time (for example, a predetermined time within a day), and the vertical axis represents the predicted value of the target data.

[0010] Graph (b) in Figure 1 shows the predicted values ​​obtained by integrating the predicted values ​​output by prediction models A1 to D1 using the technology described in Patent Document 2, and the corresponding actual values, shown as solid and dashed lines, respectively. The horizontal and vertical axes are the same as in Graph (a) in Figure 1. These predicted values ​​are the average of the predicted values ​​output by prediction models A1 to D1.

[0011] Graph (c) in Figure 1 shows the predicted values ​​obtained by integrating the predicted values ​​output by prediction models A1 to D1 using the technology described in Patent Document 3, and the corresponding actual values, shown as solid and dashed lines, respectively. The horizontal and vertical axes are the same as in Graph (a) in Figure 1. This predicted value is the one selected predicted value from among the predicted values ​​output by prediction models A1 to D1 that shows the best error trend with respect to the actual values.

[0012] As shown in Figure 1, when the predicted values ​​output by prediction models A1 to D1 are scattered without bias above or below the actual values, integrating them using the technology of Patent Document 2 tends to result in a smaller discrepancy between the predicted and actual values ​​than using the technology of Patent Document 3. Therefore, when inferring predicted values ​​of the target data at a time in the future from the current time t0, it can be expected that the prediction performance will be higher when using the technology of Patent Document 2.

[0013] On the other hand, Figure 2 shows how the predicted values ​​output by prediction models A2 to C2, which are different from the prediction models A1 to D1 shown in Figure 1, are integrated using the technology described in Patent Documents 2 and 3.

[0014] Graph (a) in Figure 2 plots the predicted values ​​output by the three prediction models A2 to C2 in the past time t0, in a time series, and connects them with a dashed line.

[0015] Graph (b) in Figure 2 shows the predicted values ​​obtained by integrating the predicted values ​​output by prediction models A2 to C2 using the technology described in Patent Document 2, and the corresponding actual values, shown as solid and dashed lines, respectively. The horizontal and vertical axes are the same as in Graph (a) in Figure 2. These predicted values ​​are the average of the predicted values ​​output by prediction models A2 to C2.

[0016] Graph (c) in Figure 2 shows the predicted values ​​obtained by integrating the predicted values ​​output by prediction models A2 to C2 using the technology described in Patent Document 3, and the corresponding actual values, shown as solid and dashed lines, respectively. The horizontal and vertical axes are the same as in Graph (a) in Figure 2. This predicted value is the one selected predicted value from among the predicted values ​​output by prediction models A2 to C2 that shows the best error trend with respect to the actual values.

[0017] As shown in Figure 2, when the predicted values ​​output by prediction models A2 to C2 are skewed upward or downward relative to the actual values, using the technology described in Patent Document 3 tends to result in a smaller discrepancy between the predicted and actual values ​​than using the technology described in Patent Document 2. Therefore, when inferring predicted values ​​for the data to be predicted at a time in the future from the current time t0, it can be expected that the prediction performance will be higher when using the technology described in Patent Document 3.

[0018] Thus, when integrating the prediction results of multiple prediction models to calculate a predicted value, depending on the degree of variability in the predicted values ​​output by each prediction model, using the technology described in Patent Document 2 may result in higher prediction performance, while using the technology described in Patent Document 3 may result in higher prediction performance. Furthermore, since the degree of variability in the predicted values ​​of each prediction model varies depending on the prediction performance of each prediction model, whether it is appropriate to use either the technology described in Patent Document 2 or 3 when integrating the prediction results of multiple prediction models to calculate a predicted value depends on the individual prediction performance of each prediction model.

[0019] For example, if the individual prediction performance of multiple prediction models varies greatly overall, the technology in Patent Document 2 is expected to improve the integrated prediction performance more than the technology in Patent Document 3 due to the ensemble effect. On the other hand, if most of the prediction models among the multiple prediction models vary relatively little, but some prediction models show prediction performance trends different from the others, the technology in Patent Document 2 may be affected by these few prediction models, potentially leading to a decrease in the integrated prediction performance. However, using the technology in Patent Document 3 avoids being affected by such a few prediction models.

[0020] For these reasons, when calculating predicted values ​​using multiple prediction models, there is a need for technologies that can reduce the influence of the prediction performance of individual prediction models and enable the calculation of predicted values ​​with greater accuracy.

[0021] The present invention has been made in view of the above-mentioned problems, and its objective is to provide a prediction system, prediction method, and program that can calculate predicted values ​​with greater accuracy by reducing the influence of the prediction performance of individual prediction models when calculating predicted values ​​using multiple prediction models. [Means for solving the problem]

[0022] A first aspect of the present invention, which solves the aforementioned problems, is a prediction system that calculates predicted values ​​of data to be predicted that change over time using a plurality of prediction models, and the predicted values ​​are derived from the values ​​of related data related to the data to be predicted. In order to find A calculation unit that calculates the calculated value for each of the prediction models using multiple prediction models that calculate the calculated value; a weight calculation unit that calculates a weight for each of the prediction models based on the prediction performance of each of the prediction models; a prediction model extraction unit that extracts one or more prediction models from the multiple prediction models whose most recent prediction performance is nearby as neighboring prediction models; the calculation value calculated for each of the neighboring prediction models by inputting the values ​​of the related data into the neighboring prediction models; and the weights of each of the neighboring prediction models. By multiplying by and adding all the multiplication results, the unified resultincluding a predicted value calculation unit that calculates the predicted value The aforementioned prediction performance refers to the degree of deviation between the calculated value or a predetermined number of the most recent calculated values ​​and the actual values ​​in a predetermined period, when continuously tracing back from the present to the past; the aforementioned neighborhood refers to a predetermined number set as the number of prediction models extracted from the plurality of prediction models in order of the highest most recent prediction performance, or a predetermined range set as the difference in prediction performance between the prediction model with the best most recent prediction performance among the plurality of prediction models. .

Advantages of the Invention

[0023] According to the present invention, when calculating a predicted value using a plurality of prediction models, it is possible to provide a prediction system, a prediction method, and a program that can reduce the influence of the prediction performance of individual prediction models and calculate the predicted value with higher accuracy.

Brief Description of the Drawings

[0024] [Figure 1] It is a diagram showing a state of integrating prediction results of a plurality of prediction models using the techniques of Patent Documents 2 and 3. [Figure 2] It is a diagram showing a state of integrating prediction results of a plurality of prediction models using the techniques of Patent Documents 2 and 3. [Figure 3] It is a block diagram showing a schematic configuration of the prediction system 1. [Figure 4A] It is a flowchart showing an example of a processing flow when the prediction model extraction unit 400 extracts the neighborhood prediction model 320(j) from the prediction models 310(i) using the neighborhood number method. [Figure 4B] It is a flowchart showing an example of a processing flow when the prediction model extraction unit 400 extracts the neighborhood prediction model 320(j) from the prediction models 310(i) using the neighborhood coefficient method. [Figure 5] It is a diagram showing an example of extracting three prediction models 310(i) with high prediction performance as the neighborhood prediction model 320(j) from four prediction models 310(i) using the neighborhood number method. [Figure 6] It is a diagram showing an example of extracting a prediction model 310(i) with high prediction performance as the neighborhood prediction model 320(j) from four prediction models 310(i) using the neighborhood coefficient method. [Figure 7] It is a block diagram showing an example of the hardware of the information processing apparatus 1000 that realizes the functions of the prediction system 1. [Figure 8]This diagram illustrates how the predicted value Ytest(t), which is integrated by the predicted value calculation unit 600, is corrected by the predicted value correction unit 800. [Modes for carrying out the invention]

[0025] The following matters will become clear from this specification and the accompanying drawings. The present invention will be described below with reference to the accompanying drawings, with reference to one embodiment thereof.

[0026] ==Prediction System 1== Figure 3 is a block diagram showing the schematic configuration of prediction system 1.

[0027] The prediction system 1 calculates predicted values ​​for data to be predicted that change over time, using multiple prediction models 310. As an example, in this embodiment, the data to be predicted is the electricity demand of a consumer, and the prediction system 1 predicts the electricity demand of a consumer every 30 minutes.

[0028] To explain in more detail, the prediction system 1 inputs relevant data that affects the electricity demand of consumers, such as weather data like temperature and humidity, and calendar data like holidays and weekdays, into multiple prediction models 310, and calculates a predicted value of electricity demand for each prediction model 310. As will be explained later, these prediction models 310 do not have the same construction methods or parameters, so even when the same relevant data is input, they output different calculated values.

[0029] The prediction system 1 then evaluates the prediction performance of each prediction model 310 by comparing the calculated values ​​for the most recent predetermined period (for example, the calculated values ​​for the most recent past 3 hours) of each prediction model 310 with the actual values ​​of electricity demand.

[0030] The prediction system 1 then calculates a weight for each prediction model 310 based on their respective prediction performance. For example, the prediction system 1 ensures that the weight value increases for prediction models 310 with relatively higher prediction performance.

[0031] Furthermore, the prediction system 1 extracts one or more prediction models 310 whose most recent prediction performance is in the vicinity as neighboring prediction models 320. For example, the prediction system 1 selects a predetermined number of prediction models 310 in descending order of their most recent prediction performance as neighboring prediction models 310, and extracts these prediction models 310 as neighboring prediction models 320.

[0032] The prediction system 1 then inputs the values ​​of the relevant data at the time to be predicted into the nearest neighbor prediction model 320, and uses the calculated values ​​calculated for each nearest neighbor prediction model 320, along with the respective weights of the nearest neighbor prediction model 320, to integrate these calculated values ​​as a predicted value (for example, by weighting the calculated values ​​by the weights).

[0033] In this configuration, when calculating predicted values ​​using multiple prediction models 310, the influence of the prediction performance of each prediction model 310 is reduced, making it possible to calculate predicted values ​​with greater accuracy.

[0034] In this embodiment, as an example, the electricity demand of a consumer every 30 minutes is used as the data to be predicted. However, various other data that change over time can also be used as the data to be predicted, such as the amount of a specific chemical substance produced in a plant in one day, or the annual harvest yield of a certain agricultural product.

[0035] As shown in Figure 3, the prediction system 1 is configured to include an input unit 100, a database 200, a prediction model construction unit 300, a calculated value unit 350, a prediction model extraction unit 400, a weight calculation unit 500, a prediction value calculation unit 600, a prediction value correction unit 800, and an output unit 700 as means for calculating predicted values ​​of the data to be predicted.

[0036] <<Input section 100>> The input unit 100 has the function of inputting or acquiring data to be predicted and related data associated with the data to be predicted from a device such as a LAN (Local Area Network) or a keyboard.

[0037] Related data are data that have some kind of correlation with the data being predicted, such as actual values, forecast values, planned values, and calculated values ​​of electricity supply and demand information (demand, supply capacity), weather information (temperature, humidity), and calendar information (holidays, weekdays).

[0038] The actual values ​​of the data to be predicted and related data are used as training data when constructing the prediction model 310. Furthermore, the forecast, planned, and calculated values ​​of the related data are input into the prediction model 310 and used to calculate the predicted values ​​of the data to be predicted.

[0039] <<Database 200>> The database 200 stores the above-mentioned data to be predicted, related data, prediction model 310, and nearest neighbor prediction model 320. In addition to the above information, the database 200 may also store data in progress during calculations within the prediction system 1.

[0040] <<Predictive Model Construction Unit 300>> The prediction model construction unit 300 constructs the above-mentioned multiple prediction models 310 using training data that associates the actual values ​​of the data to be predicted with the actual values ​​of related data.

[0041] In this embodiment, the prediction model construction unit 300 constructs M prediction models 310(i) (where i is a natural number from 1 to M) (where i is a natural number from 1 to M).

[0042] At this time, the prediction model construction unit 300 constructs each prediction model 310 using one of several types of machine learning algorithms, including (a) linear methods, (b) interval linear methods, and (c) nonlinear methods, and ensures that the M prediction models 310 as a whole contain a mixture of two or more methods. For example, the prediction model construction unit 300 includes at least two of the above (a) to (c) machine learning algorithms.

[0043] Linear methods approximate the relationship between the target variable (actual values ​​of the data to be predicted) and the explanatory variables (actual values ​​of related data) in the training data using a linear function with a straight line. Examples of machine learning algorithms that employ linear methods include linear regression models and ARIMA (Auto Regressive Integrated Moving Average) models.

[0044] Interval linear modeling is a method that divides the relationship between the target variable and explanatory variables in the training data into multiple intervals and approximates the relationship between the target variable and explanatory variables in each interval with a straight line. Machine learning algorithms that can be used for interval linear modeling include, for example, the k-nearest neighbors method and JIT (Just-In-Time) modeling.

[0045] Nonlinear methods approximate the relationship between the target variable and explanatory variables in training data using various nonlinear functions such as exponential, logarithmic, power, and trigonometric functions. Nonlinear machine learning algorithms can include neural networks, SVR (Support Vector Regression), XGBoost (eXtreme Gradient Boosting), and deep learning.

[0046] The prediction performance of each individual prediction model 310 is influenced by the machine learning algorithm used to construct the prediction model 310. However, as described above, by constructing M prediction models 310(i) using multiple machine learning algorithms with different methods, diversity in the prediction results of the M prediction models 310(i) can be ensured.

[0047] For example, even if the season, time of day, or region changes, it can be expected that at least one of the M prediction models 310(i) will exhibit high prediction performance. For instance, at a certain time or place, the prediction performance of the prediction model 310 constructed using a linear method may be higher than that of the prediction model 310 constructed using other methods. At a different time or place, the prediction performance of the prediction model 310 constructed using a nonlinear method may be higher than that of the prediction model 310 constructed using other methods. In this way, under any conditions, at least one of the M prediction models 310(i) can be expected to exhibit high prediction performance.

[0048] In this configuration, when calculating predicted values ​​using multiple prediction models 310, the influence of the prediction performance of each prediction model 310 is reduced, making it possible to calculate predicted values ​​with greater accuracy.

[0049] Furthermore, various parameters such as the training period and the number of training data points may be arbitrarily set according to the desired prediction performance for the M prediction models 310(i).

[0050] <<Calculated Value Unit 350>> The calculation unit 350 inputs the values ​​of related data into multiple prediction models 310, and calculates the predicted value of the data to be predicted for each prediction model 310.

[0051] These calculated values ​​are derived from the same related data, but they vary for each prediction model 310 due to differences in the prediction performance of each model 310.

[0052] <<Predictive Model Extraction Unit 400>> The prediction model extraction unit 400 extracts, from the M prediction models 310(i) constructed by the prediction model construction unit 300, a number of prediction models 310(i) less than or equal to the number of prediction models 310(i) (L (where L is a natural number)) as neighboring prediction models 320(j) (j = a natural number from 1 to L).

[0053] Furthermore, the prediction performance of the M prediction models 310(i) can be calculated using the most recent predetermined values ​​when continuously going back from the present to the past, or using the most recent predetermined number of calculated values.

[0054] Furthermore, the prediction performance of the prediction model 310 indicates the degree of discrepancy between the calculated values ​​and actual values ​​of the data to be predicted. The smaller the degree of discrepancy, the higher the prediction performance, and the larger the degree of discrepancy, the lower the prediction performance. The prediction performance of the M prediction models 310(i) is calculated based on the calculated values ​​obtained using the M prediction models 310(i) and the actual values ​​corresponding to these calculated values.

[0055] The prediction model extraction unit 400 can extract L neighboring prediction models 320(j) using two methods, for example, the neighboring number method and the neighboring coefficient method.

[0056] The nearest neighbor count method extracts a predetermined number of nearest neighbors (a set number) K (=L) from among M prediction models 310(i), selecting the prediction models (i) with the highest prediction performance as nearest neighbor prediction models 320(j). The nearest neighbor count method is effective when the number of prediction models that yield an ensemble effect is known. On the other hand, the nearest neighbor coefficient method extracts prediction models 310(i) from among M prediction models 310(i) such that the difference in prediction performance between the best prediction model with the highest prediction performance is within a predetermined range of nearest neighbor coefficients Kcoffe, and these are selected as nearest neighbor prediction models 320(j). The nearest neighbor coefficient method is effective when the characteristics of the prediction performance of the M prediction models 310(i) change frequently, and the number L of nearest neighbor prediction models 320(j) extracted varies depending on the characteristics of the prediction performance of the M prediction models 310(i).

[0057] Figure 4A is a flowchart illustrating an example of the processing flow when the prediction model extraction unit 400 extracts L nearest neighbor prediction models 320(j) from M prediction models 310(i) using the nearest neighbor method.

[0058] First, the prediction model extraction unit 400 calculates a numerical value representing the prediction performance from the past calculated values ​​of the data to be predicted in the M prediction models 310(i) and the corresponding actual values ​​(S1010). If the prediction performance is calculated from the past calculated values ​​and actual values ​​of the data to be predicted, for example, the mean absolute error MAE(i) or the coefficient of determination R2(i) shown in the following formula can be used. Note that R2 represents R squared. In addition to the mean absolute error MAE(i) or the coefficient of determination R2(i), the root mean square error or AIC (Akaike Information Criterion) can also be used.

[0059] The mean absolute error MAE(i) can be expressed by the following equation (1).

number

[0060] Furthermore, the coefficient of determination R²(i) can be expressed by the following equation (2).

number

[0061] Since a smaller Mean Absolute Error (MAE) value indicates higher prediction performance, the reciprocal of the Mean Absolute Error (MAE) can be used as the Index (i), a value representing the prediction performance of the prediction model (i), as shown in equation (3) below.

[0062]

number

[0063] Furthermore, since a larger R2(i) value indicates higher predictive performance, the coefficient of determination R2(i) can be used directly as Index(i), a value representing the predictive performance of the prediction model 310(i), as shown in equation (4) below.

[0064]

number

[0065] Thus, by using equation (3) or equation (4), the prediction performance when extracting the nearest neighbor prediction model 320(j) can be quantified. Furthermore, when extracting L nearest neighbor prediction models 320(j) based on the coefficient of determination R2(i) of equation (4), if the coefficient of determination R2(i) is a negative value, the value of the coefficient of determination R2(i) can be replaced with 0, and prediction models 310(i) with a coefficient of determination R2(i) value of 0 can be considered as prediction models with low prediction performance and excluded from the selection of nearest neighbor prediction models 320(j).

[0066] In step S1010, once an Index(i) value representing the prediction performance of the M prediction models 310(i) is calculated, the prediction model extraction unit 400 extracts a predetermined number of neighbors K prediction models 310(i) from the M prediction models 310(i) in descending order of Index(i) as neighbor prediction models 320(j) (S1020).

[0067] Furthermore, if a flag KM(i) is set to distinguish whether one of the M prediction models 310(i) corresponds to the nearest prediction model 320(j), this flag KM(i) can be expressed by the following equation (5). The flag KM(i) is 1 for prediction models 310(i) that correspond to the nearest prediction model 320(j) among the M prediction models 310(i), and 0 for prediction models 310(i) that do not correspond to the nearest prediction model 320(j).

[0068]

number

[0069] Figure 5 shows an example of using the nearest neighbor method to extract, for example, three prediction models 310(i) with high prediction performance from four prediction models 310(i) (i=1 to 4) as the nearest neighbor prediction models 320(j) (j=1 to 3). In Figure 5, the number of nearest neighbors K is set to 3 beforehand.

[0070] Graph (a) in Figure 5 shows an example of the calculated values ​​of four target data points inferred by four prediction models 310(i) prior to the current time t0. The horizontal axis represents time (e.g., a predetermined time within a day), and the vertical axis represents the calculated values ​​of the target data points. Since actual values ​​of the target data points also exist prior to the current time t0, the actual values ​​are shown with a dashed line and the calculated values ​​with a dotted line. Note that the three nearest neighbor prediction models 320(j) have not yet been extracted in graph (a).

[0071] The Index(i) (i=1~4), which represents the prediction performance of the four prediction models 310(i), can be obtained by equation (3) or equation (4). In Figure 5, for example, we assume that Index(1)=0.37 for prediction model 310(1), Index(2)=0.48 for prediction model 310(2), Index(3)=0.85 for prediction model 310(3), and Index(4)=0.91 for prediction model 310(4).

[0072] Graph (b) in Figure 5 shows an example of calculated values ​​for three data points selected according to the nearest neighbor method from the four data points shown in Graph (a) in Figure 5. The horizontal and vertical axes are the same as in Graph (a) in Figure 5. In the past time t0, actual values ​​of the data points to be predicted also exist, so the actual values ​​are shown as a dashed line, the calculated values ​​of prediction model 310(i) that were not extracted are shown as a dashed line, and the calculated values ​​of the extracted prediction model 310(i) are shown as a solid line.

[0073] Among the prediction models 310(1) to 310(4), the prediction models with the highest prediction performance are, as is clear from the prediction performance values ​​Index(1) to (4), prediction models 310(4), 310(3), 310(2), and 310(1), in that order. Since the number of neighbors is 3, the prediction model extraction unit 400 extracts prediction models 310(4), 310(3), and 310(2) from among the prediction models 310(1) to 310(4) as neighbor prediction models 320(1), 320(2), and 320(3), respectively.

[0074] Figure 4B is a flowchart showing an example of the processing flow when the prediction model extraction unit 400 extracts L nearest neighbor prediction models 320(j) from M prediction models 310(i) using the nearest neighbor coefficient method.

[0075] First, the prediction model extraction unit 400 calculates an Index(i) shown in equation (3) or equation (4), which is a value indicating prediction performance, from the past calculated values ​​of the data to be predicted in the M prediction models 310(i) and the actual values ​​corresponding to these calculated values, in the same manner as the process in step S1010 in Figure 4A (S2010).

[0076] Next, the prediction model extraction unit 400 extracts the prediction model 310(i) with the maximum Index(i) in equation (3) or equation (4) from among the M prediction models 310(i) as the best prediction model Mbest (S2020). The best prediction model Mbest can be expressed by the following equation (6).

[0077]

number

[0078] Next, the prediction model extraction unit 400 extracts, from among the M prediction models 310(i), the prediction model 310(i) whose Index(i) difference from the best prediction model Mbest is within a predetermined range (within the range of the neighboring coefficient Kcoffe), as the neighboring prediction model 320(j) (S2030). The neighboring coefficient Kcoffe is an arbitrarily set value, and the best prediction model Mbest is included among the neighboring prediction models 320(j).

[0079] Furthermore, if a flag KM(i) is set to distinguish whether one of the M prediction models 310(i) corresponds to the nearest prediction model 320(j), this flag KM(i) can be expressed by the following equation (7). The flag KM(i) is 1 for prediction models 310(i) that correspond to the nearest prediction model 320(j) among the M prediction models 310(i), and 0 for prediction models 310(i) that do not correspond to the nearest prediction model 320(j).

[0080]

number

[0081] Furthermore, the flag KM(i) that distinguishes whether one of the M prediction models 310(i) corresponds to a neighboring prediction model 320(j) can also be expressed as shown in equation (8) below, instead of equation (7). Equation (8) compares the ratio of the prediction performance index(i) of the other prediction models 310(i) to the prediction performance index(Mbest) of the best prediction model Mbest with 1-Kcoffe, making it possible to easily set the neighboring coefficient Kcoffe without depending on the data to be predicted, and reducing the engineering cost when setting the neighboring coefficient Kcoffe. Thus, the above predetermined range can be defined as the acceptable range of the ratio between the prediction performance of the prediction model 310(i) with the best prediction performance and the prediction performance of the prediction model 310(i) that does not have the best prediction performance.

[0082]

number

[0083] Figure 6 shows an example of using the nearest neighbor coefficient method to extract, for example, a prediction model 310(i) from among four prediction models 310(i) (i=1~4) such that the difference in Index(i) between it and the best prediction model Mbest is within the range of the nearest neighbor coefficient Kcoffe, and this model is selected as the nearest neighbor prediction model 320(j). In Figure 6, the nearest neighbor coefficient Kcoffe is set to 0.1 beforehand.

[0084] Graph (a) in Figure 6 shows an example of the calculated values ​​of four target data points inferred by four prediction models 310(i) in the past of the current time t0. The horizontal axis represents time (e.g., a predetermined time within a day), and the vertical axis represents the calculated values ​​of the target data points. Since actual values ​​also exist in the past of the current time t0, the actual values ​​are shown as a dashed line, and the calculated values ​​of prediction model 310(i) are shown as a dashed line. Note that the nearest neighbor prediction model 320(j) has not yet been extracted in Graph (a) in Figure 6.

[0085] The Index(i) (i=1~4), which represents the prediction performance of the four prediction models 310(i), is obtained by equation (3) or equation (4). In Figure 6, as in Figure 5, we assume that Index(1)=0.37 for prediction model 310(1), Index(2)=0.48 for prediction model 310(2), Index(3)=0.85 for prediction model 310(3), and Index(4)=0.91 for prediction model 310(4).

[0086] Graph (b) in Figure 6 shows an example of selecting the calculated values ​​from the best prediction model Mbest among the calculated values ​​of the four data sets to be predicted shown in Graph (a) of Figure 6. The horizontal and vertical axes are the same as in Graph (a) of Figure 6. Since actual values ​​also exist for periods prior to the current time t0, the actual values ​​are shown as a dashed line, the calculated values ​​of the best prediction model Mbest as a solid line, and the calculated values ​​of prediction models 310(i) other than the best prediction model Mbest as a dashed line.

[0087] Graph (c) in Figure 6 shows an example of selecting calculated values ​​from the four data points to be predicted shown in Graph (a) of Figure 6, where the difference from the calculated value in the best prediction model Mbest is within the range of the nearest neighbor coefficient Kcoffe. The horizontal and vertical axes are the same as in Graph (a) of Figure 6. In the nearest neighbor coefficient method, two prediction models 310(4) and 310(3) are extracted as the nearest neighbor prediction model 320(j).

[0088] Furthermore, depending on the value of Index(i) and the neighbor coefficient Kcoffe of prediction model 310(i), the number of neighbor prediction models 320(j) extracted, L, may be as low as one.

[0089] <<Weight calculation unit 500>> The weight calculation unit 500 calculates the weight of each prediction model 310(i) based on the prediction performance of each prediction model 310(i). For example, the weight calculation unit 500 uses a value Index(i) indicating the prediction performance of the prediction model 310(i) and a flag KM(i) that distinguishes whether or not it corresponds to a neighboring prediction model 320(j), to calculate the weights W(i) of the M prediction models 310(i) according to the following equation (9). Note that, as explained below, if the weight W(i) of a prediction model 310(i) that was not extracted as a neighboring prediction model 320(j) is set to 0, then the weight W(i) in equation (9) can also be read as the ratio of the sum of the values ​​obtained by multiplying the prediction performance Index of the extracted neighboring prediction models 320(j) by the flag KM, to the value obtained by multiplying the corresponding one neighboring prediction model 320(j) by the flag KM.

[0090]

number

[0091] Here, the values ​​Index(i) and Index(m) (m=1~M) shown in the numerator and denominator of equation (9), which indicate the prediction performance of the prediction model 310(i), are either equation (3), which uses the reciprocal of the mean absolute error MAE(i), or equation (4), which uses the coefficient of determination R2(i) as is. Also, the flags KM(i) and KM(m) (m=1~M) shown in the numerator and denominator are either equation (5), which is used in the nearest neighbor number method, or equations (7) and (8), which are used in the nearest neighbor coefficient method.

[0092] The numerator of equation (9) shows a term that multiplies the prediction performance of the prediction model 310(i) for which the weight W(i) is to be calculated by the value Index(i) and the flag KM(i). The denominator of equation (9) shows a term that multiplies the prediction performance of each of the M prediction models 310(i) by the value Index(m) and the flag KM(m), and then adds up all the multiplication results.

[0093] According to equation (9), the weight W(i) of the prediction model 310(i) that was not selected as the nearest neighbor prediction model 320(j) is 0. Also, the weight W(i) of each prediction model 310(i) that was selected as the nearest neighbor prediction model 320(i) is calculated such that the sum of these weights W(i) is 1.

[0094] <<Prediction Value Calculation Unit 600>> The prediction value calculation unit 600 calculates the predicted value of the data to be predicted by integrating the calculated values ​​obtained for each of the L nearest neighbor prediction models 320(j) by inputting the values ​​of the relevant data at the time to be predicted into the L nearest neighbor prediction models 320(j), and the respective weights W(j) of the nearest neighbor prediction models 320(j).

[0095] In this embodiment, the prediction value calculation unit 600 calculates the prediction value Ytest(t) by integrating the calculated values ​​Xtest(t,i) of the prediction target data from M prediction models 310(i) at time t using a weighted average of the calculated values ​​Xtest(t,i) and weights W(i). Specifically, the prediction value calculation unit 600 calculates the integrated prediction value Ytest(t) by multiplying the calculated values ​​Xtest(t,i) of the prediction target data from each of the M prediction models 310(i) by the respective weights W(i) of the M prediction models 310(i) according to the following equation (10), and then adding up all the multiplication results.

[0096]

number

[0097] Since the weights W(i) of the prediction models 310(i) that were not selected as neighboring prediction models 320(j) from among the M prediction models 310(i) are 0, the integrated prediction value Ytest(t) is calculated by the weighted average of the calculated value Xtest(t,i) of the prediction models 310(i) selected as neighboring prediction models 320(j) from among the M prediction models 310(i) and their weights W(i), and is a value that does not depend on the prediction performance of the M prediction models 310(i).

[0098] <<Prediction value correction unit 800>> The predicted value Ytest(t) calculated as described above is calculated using L neighbor prediction models 320(j) and weights W(j) which have relatively high prediction performance, and is therefore expected to have reasonably high accuracy. However, it is still difficult to make the difference from the actual value zero.

[0099] Therefore, the prediction value correction unit 800 corrects the predicted value Ytest(t) based on the most recent actual value. Specifically, the prediction value correction unit 800 corrects the future predicted value Ytest(t) from the present time onward using the difference between the predicted value Ytest(t) and the actual value for a predetermined period in the past up to the present time.

[0100] Figure 8 illustrates how the predicted value Ytest(t) calculated by the predicted value calculation unit 600 is corrected by the predicted value correction unit 800.

[0101] Graph (a) in Figure 8 shows the predicted value Ytest(t) before correction by the prediction value correction unit 800 and the actual value for a predetermined period in the past up to the present time. The horizontal axis represents time (for example, a predetermined time within a day), and the vertical axis represents the predicted value Ytest(t) and the actual value. The predicted value Ytest(t) before correction is shown as a dashed line, and the actual value is shown as a dashed line. The left end of the predicted value Ytest(t) before correction is the time a predetermined time before the present time (hereinafter referred to as the difference calculation start time), and the right end of the actual value is the present time.

[0102] JPEG0007835334000012.jpg31170

[0103] Furthermore, K is the number of predicted values ​​Ytest(t) that exist between the start time of the difference calculation and the current time. Also, T is the number of predicted values ​​Ytest(t) that the predicted value correction unit 800 corrects after the current time.

[0104] JPEG0007835334000013.jpg37170

[0105]

number

[0106] The operation of the predicted value correction unit 800 will be explained below with reference to Figure 8.

[0107] JPEG0007835334000016.jpg26170

[0108] JPEG0007835334000017.jpg32170

[0109] In this configuration, even if the trend of recent actual values ​​of the data to be predicted changes abruptly for some reason, causing a discrepancy between the predicted and actual values, the predicted values ​​can be immediately corrected to eliminate the discrepancy, and the corrected predicted values ​​can be smoothly continued from the uncorrected predicted values.

[0110] TIFF0007835334000018.tif15170

[0111] ==Information Processing Device 1000== Figure 7 is a block diagram showing an example of the hardware of the information processing device 1000 that realizes the functions of the prediction system 1.

[0112] The information processing device 1000 comprises a processor 1010, a main memory 1020, an auxiliary memory 1030, an input device 1040, an output device 1050, and a communication device 1060.

[0113] The information processing device 1000 may be, for example, a personal computer, an office computer, various server devices, a general-purpose computer, etc. The information processing device 1000 may be implemented in whole or in part using virtual information processing resources provided using virtualization technology, such as a virtual server provided by a cloud system.

[0114] The processor 1010 is composed of components such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit), and AI (Artificial Intelligence) chip.

[0115] The main memory 1020 is a device for storing programs and data, and can be, for example, ROM (Read Only Memory), RAM (Random Access Memory), or non-volatile memory (NVRAM (Non-Volatile RAM)).

[0116] The auxiliary storage device 1030 is, for example, an SSD (Solid State Drive), a hard disk drive, an optical storage device (CD (Compact Disc), DVD (Digital Versatile Disc), etc.), a storage system, an IC card, an SD card, a reader / writer for optical recording media, or the storage area of ​​a cloud server. Programs and data can be read into the auxiliary storage device 1030 via a recording media reader or a communication device 1060. Programs and data stored in the auxiliary storage device 1030 are read into the main memory 1020 as needed.

[0117] The input device 1040 is an interface that accepts input from an external source, and can be, for example, a keyboard, mouse, touch panel, card reader, pen-input tablet, or voice input device.

[0118] The output device 1050 is an interface that outputs various information such as processing progress and processing results. The output device 1050 is, for example, a display device that visualizes the above information (LCD (Liquid Crystal Display), graphics card, etc.), a device that converts the above information into sound (speaker, etc.), or a device that converts the above information into text (printer, etc.). The information processing device 1000 may also be configured to input and output information to and from other devices via the communication device 1060.

[0119] The input device 1040 and the output device 1050 constitute a user interface for receiving and presenting information with the user.

[0120] The communication device 1060 is a device that enables communication (wired or wireless communication) with other devices via a communication infrastructure, and is configured using, for example, a NIC (Network Interface Card), a wireless communication module, a USB module, etc.

[0121] Furthermore, the information processing device 1000 may have, for example, an operating system, a file system, a DBMS (Database Management System) (relational database, NoSQL, etc.), a KVS (Key-Value Store), etc. installed on it.

[0122] The functions of the prediction model construction unit 300, the calculated value calculation unit 350, the prediction model extraction unit 400, the weight calculation unit 500, the predicted value calculation unit 600, and the predicted value correction unit 800 are realized by the processor 1010 of the information processing device 1000 reading and executing control programs and various data stored in the main memory 1020 and the auxiliary memory 1030. The database 200 is realized by the main memory 1020 and the auxiliary memory 1030. The input unit 100 is realized by the input device 1040 and the communication device 1060, and the output unit 700 is realized by the output device 1050 and the communication device 1060.

[0123] ==Summary== As described above, the prediction system 1 according to this embodiment is a prediction system 1 that calculates predicted values ​​of data to be predicted that change over time using a plurality of prediction models 310, and includes: a calculation value calculation unit 350 that calculates the calculated value for each prediction model 310 using a plurality of prediction models 310 that calculate the calculated value from the values ​​of related data related to the data to be predicted; a weight calculation unit 500 that calculates a weight for each prediction model 310 based on the prediction performance of each prediction model 310; a prediction model extraction unit 400 that extracts one or more prediction models 310 from the plurality of prediction models 310 whose most recent prediction performance is in the vicinity as neighboring prediction models 320; and a prediction value calculation unit 600 that calculates the predicted value by integrating the calculated value calculated for each neighboring prediction model 320 using the calculated value calculated for each neighboring prediction model 320 by inputting the values ​​of the related data into the neighboring prediction model 320, and the respective weights of the neighboring prediction model 320.

[0124] Furthermore, the prediction system 1 includes a prediction value correction unit 800 that corrects the prediction value based on the most recent actual value of the data to be predicted.

[0125] Furthermore, in the prediction system 1, the prediction value correction unit 800 corrects the prediction values ​​from the current time onward based on the difference between a plurality of prediction values ​​calculated in time series within the most recent predetermined period and the actual value corresponding to each prediction value.

[0126] Furthermore, the prediction system 1 further includes a prediction model construction unit 300 that constructs the plurality of prediction models 310 based on multiple types of machine learning algorithms.

[0127] Furthermore, in prediction system 1, Multiple types of machine learning algorithms include at least two of the following: linear methods, piecewise linear methods, and nonlinear methods.

[0128] Furthermore, in prediction system 1, the prediction performance of prediction model 310 is calculated based on the calculated value obtained using prediction model 310 and the actual value corresponding to this calculated value.

[0129] Furthermore, in the prediction system 1, the prediction model extraction unit 400 extracts the nearest neighbor prediction model 320 based on the relationship between the prediction performance of all M prediction models 310.

[0130] Furthermore, in the prediction system 1, the prediction model extraction unit 400 selects a predetermined number of prediction models 310 from among the M prediction models 310 in order of their most recent prediction performance, as the nearest prediction models 310 in terms of their most recent prediction performance, and extracts these prediction models 310 as the nearest prediction models 320.

[0131] Furthermore, in the prediction system 1, the prediction model extraction unit 400 selects from among the M prediction models 310 a prediction model 310 whose prediction performance difference from the prediction model 310 with the best most recent prediction performance is within a predetermined range, and extracts this prediction model 310 as the nearest prediction model 320.

[0132] Furthermore, in the prediction system 1, the above-mentioned predetermined range is defined as the acceptable range of the ratio between the prediction performance of the prediction model 310 that has the best prediction performance and the prediction performance of the prediction model 310 that does not have the best prediction performance.

[0133] Furthermore, in the prediction system 1, the weight calculation unit 500 sets the weights of the prediction models 310 that have not been extracted as neighboring prediction models 320 out of the M prediction models 310 to 0.

[0134] Furthermore, in the prediction system 1, the weight calculation unit 500 calculates weights such that the sum of the weights of the neighboring prediction model 320 equals 1.

[0135] Furthermore, in the prediction system 1, the weight calculation unit 500 calculates weights based on the ratio of the prediction performance of each of the L nearest neighbor prediction models 320.

[0136] Furthermore, in the prediction system 1, the prediction value calculation unit 600 integrates the calculated values, which are calculated for each of the neighboring prediction models 320, into the prediction value by weighting them with the weights corresponding to the neighboring prediction model 320 that calculated each calculated value.

[0137] Furthermore, by employing prediction system 1, when calculating predicted values ​​using multiple prediction models 310, the influence of the prediction performance of each individual prediction model 310 is reduced, making it possible to calculate predicted values ​​with greater accuracy.

[0138] This embodiment is provided to facilitate understanding of the present invention and is not intended to limit its interpretation. The present invention may be modified or improved without departing from its spirit, and equivalents thereof are also included.

[0139] For example, the processing flow shown in Figure 4A, which extracts the nearest neighbor prediction model 320(j) using the nearest neighbor count method, and the processing flow shown in Figure 4B, which extracts the nearest neighbor prediction model 320(j) using the nearest neighbor coefficient method, can be executed in parallel. The processing from Figure 4A to Figure 4B that extracts the nearest neighbor prediction model 320(j) with higher prediction performance can then be used for subsequent processing. [Explanation of symbols]

[0140] 1 Prediction system, 100 Input unit, 200 Database, 300 Prediction model construction unit, 350 Calculated value unit, 400 Prediction model extraction unit, 500 Weight calculation unit, 600 Predicted value calculation unit, 700 Output unit, 800 Predicted value correction unit, 1000 Information processing unit, 1010 Processor, 1020 Main memory, 1030 Auxiliary memory, 1040 Input device, 1050 Output device, 1060 Communication device

Claims

1. A prediction system that calculates predicted values ​​of data to be predicted that change over time using multiple prediction models, A calculation value calculation unit that calculates the calculated value for each of the prediction models, using a plurality of prediction models that calculate calculated values ​​for obtaining the predicted value from the values ​​of related data related to the data to be predicted, A weight calculation unit calculates weights for each of the prediction models based on the prediction performance of each of the prediction models, A prediction model extraction unit extracts one or more prediction models from the aforementioned multiple prediction models, each of which has the most recent prediction performance in the vicinity, as neighboring prediction models. The system includes a prediction value calculation unit that calculates an integrated prediction value by multiplying the calculated value, which is calculated for each of the neighboring prediction models by inputting the values ​​of the related data into the neighboring prediction model, by the respective weights of the neighboring prediction models, and then adding up all the multiplication results. The aforementioned prediction performance refers to the degree of deviation between the calculated value or a predetermined number of the most recent calculated values ​​and the actual values ​​when continuously tracing back from the present to the past. The term "neighborhood" refers to a predetermined number of prediction models selected from the plurality of prediction models in order of their most recent prediction performance, or a predetermined range defined as the difference in prediction performance between the prediction model with the best recent prediction performance among the plurality of prediction models. Prediction system.

2. A prediction system according to claim 1, Prediction value correction unit corrects the aforementioned predicted value based on the most recent actual value of the data to be predicted. A prediction system that includes this.

3. A prediction system according to claim 2, The aforementioned forecast value correction unit corrects the forecast values ​​from the current time onward based on the difference between multiple forecast values ​​calculated in time series within the most recent predetermined period and the actual values ​​corresponding to each forecast value. Prediction system.

4. A prediction system according to claim 1, Predictive model construction unit that constructs the aforementioned multiple predictive models based on multiple types of machine learning algorithms A prediction system that includes this.

5. A prediction system according to claim 4, The aforementioned machine learning algorithms of multiple types include at least two of the following: linear methods, piecewise linear methods, and nonlinear methods. Prediction system.

6. A prediction system according to claim 1, The prediction model extraction unit extracts the nearest neighbor prediction model based on the relationship between the prediction performance of all the multiple prediction models. Prediction system.

7. A prediction system according to claim 1, The predetermined range is defined as the acceptable range for the ratio between the prediction performance of the prediction model with the best prediction performance and the prediction performance of the prediction model that does not have the best prediction performance. Prediction system.

8. A prediction system according to claim 1, The weight calculation unit sets the weights of prediction models that have not been extracted as the nearest neighbor prediction models from among the plurality of prediction models to 0. Prediction system.

9. A prediction system according to claim 8, The weight calculation unit calculates the weights such that the sum of the weights of the nearest neighbor prediction model equals 1. Prediction system.

10. A prediction system according to claim 1, The weight calculation unit calculates the weights based on the ratio of the prediction performance of each of the neighboring prediction models. Prediction system.

11. A prediction method for an information processing device that calculates predicted values ​​of data to be predicted that change over time using multiple prediction models, A calculation value calculation step, which involves calculating the calculated value for each prediction model using a plurality of prediction models that calculate calculated values ​​for obtaining the predicted value from the values ​​of related data related to the data to be predicted, A weight calculation step in which a weight is calculated for each of the prediction models based on the prediction performance of each of the prediction models, From the aforementioned multiple prediction models, one or more prediction models whose most recent prediction performance is in the vicinity are selected as neighboring prediction models in an extraction step. The prediction value calculation step includes multiplying the calculated value obtained for each nearest neighbor prediction model by inputting the values ​​of the related data into the nearest neighbor prediction model by the respective weights of the nearest neighbor prediction model, and then adding up all the multiplication results to calculate the integrated prediction value, The aforementioned prediction performance refers to the degree of deviation between the calculated value or a predetermined number of the most recent calculated values ​​and the actual values ​​when continuously tracing back from the present to the past. The term "neighborhood" refers to a predetermined number of prediction models selected from the plurality of prediction models in order of their most recent prediction performance, or a predetermined range defined as the difference in prediction performance between the prediction model with the best recent prediction performance among the plurality of prediction models. Prediction method.

12. A program that causes a computer to calculate predicted values ​​for data to be predicted that change over time, using multiple prediction models, To the aforementioned computer, A calculation value calculation process that calculates the calculated value for each of the prediction models, using a plurality of prediction models that calculate calculated values ​​for obtaining the predicted value from the values ​​of related data related to the data to be predicted, A weight calculation process is performed to calculate weights for each of the prediction models based on their respective prediction performance. An extraction process is performed to extract one or more prediction models from the aforementioned multiple prediction models, each of which has the most recent prediction performance in the vicinity, as neighboring prediction models. The system performs a prediction value calculation process that calculates an integrated prediction value by inputting the values ​​of the related data into the nearest neighbor prediction model, multiplying the calculated value calculated for each nearest neighbor prediction model by the respective weights of the nearest neighbor prediction model, and then adding up all the multiplication results. The aforementioned prediction performance refers to the degree of deviation between the calculated value or a predetermined number of the most recent calculated values ​​and the actual values ​​when continuously tracing back from the present to the past. The term "neighborhood" refers to a predetermined number of prediction models selected from the plurality of prediction models in order of their most recent prediction performance, or a predetermined range defined as the difference in prediction performance between the prediction model with the best recent prediction performance among the plurality of prediction models. program.

Images