Hybrid Forecasting Model for Time-Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face decreased forecasting performance due to correlations between explanatory variables and the presence of noise, and physical models struggle with accurate long-term forecasting and consideration of external control information.
Innovation Solution
An information processing apparatus that fuses a physical model and a machine learning model to construct a forecasting model, using time-series data to calculate forecasting residuals and improve prediction accuracy by integrating both models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is used for forecasting, then forecasting accuracy is improved by handling control data, but forecasting performance deteriorates due to correlations between explanatory variables and noise
Solution Approach 1:
The patent segments the forecasting model into two distinct components: a physical model that handles the deterministic relationships and control data, and a machine learning model that captures residual patterns. This segmentation allows each model to specialize in what it does best, avoiding the pitfalls of using a single model type for all forecasting tasks.
Solution Approach 2:
The patent merges the physical model and machine learning model into a hybrid forecasting system. The physical model provides baseline forecasts based on control data and physical relationships, while the machine learning model learns from the residuals to capture complex nonlinear patterns, achieving synergistic improvement in both accuracy and reliability.
2Duration of action of moving object
If a physical model is used for forecasting, then long-term forecasting capability is improved, but ability to handle control data and external factors is reduced
Solution Approach 1:
The patent segments the forecasting functionality between two models: the physical model handles long-term trends and deterministic behavior, while the machine learning model handles short-term fluctuations and control data effects. This segmentation enables the system to leverage the strengths of both approaches across different time horizons.
Solution Approach 2:
The hybrid forecasting system achieves multi-functionality by enabling the physical model to provide stable long-term forecasts while the machine learning model simultaneously captures the effects of control data and external factors. The combination creates a universal system that can handle both long-duration forecasting and adaptive response to control inputs.
3Loss of information
If explanatory variables with correlations are used in machine learning, then comprehensive information is captured, but forecasting performance deteriorates
Solution Approach 1:
The patent extracts the deterministic relationships and control data effects into the physical model component, separating these from the machine learning model. This extraction allows the machine learning model to focus on learning residual patterns without being confounded by correlated explanatory variables, thereby maintaining forecasting accuracy while still capturing comprehensive information.
4Duration of action of moving object
If the forecasting period is extended in a machine learning model, then long-term prediction capability is achieved, but forecasting error increases
Solution Approach 1:
The patent segments the forecasting task by using the physical model to provide stable long-term forecasts where deterministic relationships dominate, while the machine learning model handles shorter-term residual patterns. This segmentation prevents the accumulation of errors that would occur if a machine learning model attempted to predict far into the future alone.
Data Source
AI summary
An information processing apparatus comprising processing circuitry, the processing circuitry inputs first learning data including time-series data to a first model and calculate a forecasted value of a target variable that is a forecasting target, calculates a first forecasting residual amount that is a deviation of the forecasted value by using second learning data; and constructs a second model for predicting the first forecasting residual amount by machine learning based on the second learning data and the first forecasting residual amount.