Energy Trading & Grid Operations

Legacy Models Transformed into a 95%+ Accurate, Self Retraining Forecasting Engine

Results delivered

95%+
target forecast accuracy hit through automated competition between LSTM, GRU, gradient boosted trees, and transformer architectures, well above the 80–85% ceiling of legacy statistical methods.
24 Hrs → 15 Days
forecast horizon range now spans real time day ahead curves out to 15 day forward schedules, covering both grid load and commodity pricing needs.
7 Year Data Lake
historical ERCOT, PJM, and natural gas market data made instantly queryable, with no manual scraping or custom ETL scripts required.

Executive Summary

SORA yHat was built to close the gap between energy market volatility and production grade forecasting. Grid operators, generation asset managers, and commodity trading desks needed a way to move beyond single method statistical models and notebook bound experiments that never made it into reliable, scheduled production. yHat automatically transforms and normalizes raw market and grid data across various AI models before training, so every architecture works from a clean, consistent feature set. Its model tournament approach then trains multiple architectures in parallel, ranks them automatically by MAPE and WAPE, and promotes the top performer to a recurring inference schedule, with no manual deployment step.

Industry

Energy & Commodity Trading

Project Type

AI Native Time Series Forecasting Platform

Services

Deep Learning Model Development · Automated Model Selection · MLOps Pipeline Design · Historical Data Lake Integration · Schema Drift Detection

Outcome

ISO power load forecasting across ERCOT and PJM now runs at hourly granularity, with peak hour predictions and load curves benchmarked against actuals in real time. On the commodity side, natural gas price forecasting runs as a standing weekly model tournament: the current production model (an XGBoost LSTM ensemble) holds a 96.4% score against a 2.3% MAPE, automatically re evaluated against 11 competing architectures. When live error rates breach a configured tolerance threshold, the platform merges new historical data and retrains without anyone triggering it manually. Teams bringing their own data (IoT telemetry, manufacturing throughput, demand logs) get the same automated schema validation and self healing retraining, turning raw CSV or S3 uploads into standing forecasting pipelines.