Probabilistic Price Spike Forecasting via Segmented Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional price forecasting systems fail to accurately predict price spikes in commodities like electricity due to their reliance on cyclical patterns, which are not followed by volatile price spikes, making it difficult for buyers and sellers to manage risk in real-time pricing environments.
Innovation Solution
A method and system that use machine learning algorithms, specifically random forest and quantile regression models, to forecast price spikes and probabilistic price predictions based on historical load, price, and time/date/location features, allowing for the identification of spike and non-spike prices and providing risk tolerance and spike level thresholds for better decision-making.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional price forecasting systems rely on cyclical patterns, then they can provide stable predictions for regular price movements, but they fail to accurately predict volatile price spikes
Solution Approach 1:
The patent segments the price forecasting problem into two distinct parts: spike detection (using a first machine learning model to identify probability of spikes) and price prediction (using a second machine learning model to predict prices when no spike occurs). This segmentation allows each model to specialize in its respective task, improving overall accuracy for both volatile and stable price movements.
Solution Approach 2:
The system dynamically switches between different forecasting approaches based on real-time conditions. When the first model detects a high probability of a price spike, the system recognizes this dynamic change and adjusts its prediction behavior accordingly, preventing inaccurate cyclical pattern predictions during volatile periods.
2Measurement precision
If a single machine learning model is used for price forecasting, then the system complexity is reduced, but the accuracy for distinguishing between spike and non-spike prices deteriorates
Solution Approach 1:
The forecasting system is segmented into two specialized models: a spike detection model that identifies probability of price spikes, and a price prediction model that forecasts prices during non-spike periods. This division of labor improves measurement precision for spike identification while keeping each individual model relatively simple.
Solution Approach 2:
The first machine learning model acts as an intermediary between raw price data and the second prediction model. It filters and classifies price movements, directing spike probabilities to the appropriate handling process, which improves overall accuracy without requiring either model to be overly complex.
3Loss of time
If price forecasting focuses on long-term trends, then investment and maintenance decisions are supported, but real-time operational decisions lack accurate short-term predictions
Solution Approach 1:
The system provides dynamic forecasting at multiple time horizons. The spike detection model operates in real-time to identify immediate price spike risks, while the price prediction model provides short-term forecasts for operational decisions. This dynamic multi-timescale approach ensures both real-time responsiveness and accurate short-term predictions.
Data Source
AI summary
A method to forecast a spike price of a commodity may include obtaining a request to perform a prediction of a spike in a price of a commodity. The method may also include obtaining additional information of the commodity based on the input. The method may also include determining, using a first machine learning algorithm, a probability of the spike based at least in part on the additional information. The method may include outputting the probability of the spike for visual representation on an electronic display in response to a determination that the probability of the spike is above a risk tolerance. The method may further include determining, using a second machine learning algorithm, a price prediction in response to a determination that the probability of the spike is below the risk tolerance, wherein determining the price prediction comprises outputting the price prediction for visual representation on the electronic display.