Machine-Learning Architecture for End-to-End Quantile Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning approaches for energy and ancillary services forecasting fail to accurately model non-stationarity, seasonality, and external factors, and are limited to deterministic point-based forecasts, lacking flexibility and failing to address market uncertainties.
Innovation Solution
A machine-learning architecture comprising a neural network stack that aggregates incremental forecast values with a seed value to produce end-to-end probabilistic forecasts, capable of modeling multiple factors without assuming a distribution, and generating quantile forecasts across different quantiles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If separate independent ML models are trained for each quantile level using quantile loss function, then probabilistic forecasting capability is achieved, but model complexity increases and quantile crossing occurs
Solution Approach 1:
The patent combines multiple quantile prediction capabilities into a single unified neural network model. The model uses a shared backbone network that processes input features once, then branches into multiple output heads that simultaneously predict different quantile levels. This merging approach eliminates the need for training separate independent models for each quantile, reducing overall model complexity while maintaining probabilistic forecasting capability.
Solution Approach 2:
The unified neural network model serves multiple functions by predicting the entire probability distribution through simultaneous quantile predictions. A single model structure handles all quantile levels (e.g., 0.1, 0.3, 0.5, 0.7, 0.9) concurrently, making the model universal rather than requiring specialized separate models for each quantile level.
2Ease of operation
If point-based deterministic forecasts are generated, then forecasting simplicity is maintained, but market uncertainty awareness is insufficient
Solution Approach 1:
The patent transitions from one-dimensional point forecasts to a multi-dimensional probability distribution representation. Instead of predicting a single deterministic value, the model predicts multiple quantile levels that collectively form a probability distribution. This adds dimensional information about uncertainty and variability, enabling traders to assess risk while maintaining the simplicity of using a single model structure.
3Device complexity
If parametric techniques are used for probabilistic forecasting, then modeling framework is simplified, but assumptions about underlying probability model are required which are not valid in energy markets
Solution Approach 1:
The patent uses flexible, data-driven quantile functions without requiring rigid parametric assumptions about the underlying probability distribution. Instead of committing to a specific parametric form (e.g., normal, log-normal), the model learns the probability distribution directly from data using quantile predictions, adapting to the actual market conditions in energy and ancillary services which may not follow standard distributions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
State-of-the-art machine-learning models for forecasting fail to address the non-stationarity and uncertainty in data, rely on assumptions about data distribution, and/or produce unusable probability distributions. Accordingly, a machine-learning architecture for end-to-end probabilistic forecasting is disclosed to address these and other problems. In particular, the machine-learning model may utilize a persistence module that outputs a seed forecast value of a target variable, a neural-network stack that predicts incremental forecast value(s) of the target variable (e.g., using back-casting), and an aggregator that aggregates the seed forecast value and the incremental forecast value(s) to produce an aggregate forecast value of the target variable. In an embodiment, this aggregate forecast value may be input to an incremental quantile module that comprises a second neural-network stack to predict the forecast value of the target variable for each of a plurality of quantiles, and which aggregates these forecast values into a probability distribution.