Differential Machine Learning Architecture for Non-Stationary Energy Price Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning models for forecasting energy prices face accuracy issues due to non-stationary data, requiring frequent retraining or increased complexity to capture long-term trends and periodicities.
Innovation Solution
A differential-based machine-learning architecture that determines differences between past and forecasted values of features and inputs these differences into a machine-learning model trained to predict differential information, allowing for accurate forecasting without frequent retraining or complex models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the machine-learning model is frequently retrained on new training data to maintain accuracy, then the forecasting accuracy is maintained, but the computational resources required increase significantly
Solution Approach 1:
The patent segments the forecasting problem into two parts: a differential model that captures short-term variations and a trend component that handles long-term patterns. This segmentation allows the model to maintain accuracy without frequent retraining, as each component can be updated independently and more efficiently.
Solution Approach 2:
The patent applies preliminary differencing to the time series data before feeding it to the machine-learning model. This preprocessing step removes trends and seasonalities in advance, allowing the model to focus on capturing short-term fluctuations, thereby maintaining accuracy with less computational effort and without frequent retraining.
2Measurement precision
If the machine-learning model is made more complex to capture long-term trends and periodicities, then the forecasting accuracy is improved, but the computational resources required to train and operate the model increase
Solution Approach 1:
The patent divides the complex task of capturing all patterns into separate components: a differential model for short-term dynamics and a separate handling of trend/seasonality. This segmentation reduces the complexity of the machine-learning model while maintaining comprehensive pattern capture through the combination of components.
Solution Approach 2:
The patent extracts and removes trends and seasonalities from the time series data before input to the machine-learning model. By taking out these long-term patterns, the model becomes simpler as it only needs to learn short-term fluctuations, yet the overall forecasting system remains accurate by reconstructing the full signal from its components.
3Adaptability or versatility
If the statistical properties of operational data shift from training data, then the machine-learning model was trained on stationary data but operational data exhibits non-stationary behavior, but the forecasting accuracy decreases over time
Solution Approach 1:
The patent applies differencing as a preliminary action to transform non-stationary data into stationary data before model input. This preprocessing ensures that the statistical properties remain stable over time, allowing the model to maintain accuracy even when operational data shifts from training data conditions.
Solution Approach 2:
The patent uses a dynamic differencing approach where the model learns the appropriate differencing order and parameters adaptively. This allows the system to handle changing data characteristics over time while maintaining stationarity assumptions, thereby preserving forecasting accuracy in non-stationary environments.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Conventional forecast models assume that the underlying data are stationary. In the case of non-stationary data, the conventional forecast model must be increased in complexity, or otherwise, must be frequently retrained. Embodiments are disclosed of a forecast model that accounts for non-stationary data, without increased complexity and without an increase in the frequency of retraining, by recasting the forecasting problem as a matter of tracking differentials in the data. In particular, variables (e.g., covariates) are input into differential determinations and/or intermediate machine-learning models to determine differences in past and forecasted values of a plurality of features. These differences are then input into a machine-learning model to predict a change in the value of the target, which is aggregated with a past value of the target to produce a forecasted value of the target.