Transformer-LSTM Hybrid Forecast Engine for RE + Storage Dispatch

Course Description

Transformer-LSTM Hybrid Forecast Engine for RE + Storage Dispatch is a three-day, hands-on program that takes you from raw data to real operational decisions. You’ll build clean, time-aligned SCADA and weather datasets, then train a late-fusion hybrid model where a Transformer learns exogenous weather patterns while an LSTM captures plant-history dynamics. The goal is practical, calibrated multi-horizon probabilistic forecasting (P10/P50/P90) that you can trust. You’ll evaluate performance using rolling backtests, document results in a concise model card, and understand what’s driving error across different regimes.

In the final stage, you’ll turn forecasts into optimal battery dispatch using rolling-horizon MPC under real operating constraints. You’ll generate dispatch and SoC trajectories and build a KPI dashboard that makes the value visible—cost savings, reserve compliance, curtailment avoided, and value-of-forecast (VFF) metrics.

Aim

To enable participants to build and deploy a Transformer–LSTM hybrid that delivers calibrated multi-horizon renewable energy forecasts (P10/P50/P90) and converts them into optimal battery dispatch through rolling-horizon MPC, with measurable KPI improvements.

Course Objectives

Participants will be able to:

• Build time-aligned SCADA and weather datasets using time-aware splits that are actually suitable for forecasting.

• Engineer practical forecasting features and establish strong baselines (persistence and LSTM) for comparison.

• Train a late-fusion Transformer–LSTM model with probabilistic multi-horizon outputs (P10/P50/P90).

• Apply training best practices, handle data gaps, and calibrate predictive uncertainty so quantiles are meaningful.

• Evaluate models using rolling-origin backtests and document results in a compact, decision-friendly model card.

• Convert probabilistic forecasts into MPC-based battery dispatch under real operational constraints.

• Plan operational deployment including serving, drift monitoring, shadow runs, and operator-facing HMI integration.

• Track and interpret dispatch and forecast KPIs, including cost savings, reserve compliance, curtailment avoided, and VFF.

Course Structure

Module 1: Data Preparation and Hybrid Forecasting Foundations

• Signals and horizons: solar/wind generation, net load, price; intraday and day-ahead forecasting use cases

• Time-aware data handling: splits, leakage avoidance, and alignment of SCADA and weather feeds

• Feature engineering: lags/rolling windows, weather look-ahead, plant metadata, and calendar effects

• Metrics and baselines: RMSE, sMAPE, multi-horizon evaluation, pinball loss; persistence baseline design

• Hybrid modeling concept: Transformer for exogenous weather signals + LSTM for plant history with late fusion

• Practical session: build a time-aligned SCADA+weather dataset; train persistence and single-layer LSTM baselines

Module 2: Training, Calibration, and Backtesting the Transformer–LSTM

• Model architecture: sequence lengths, attention encoder–decoder, LSTM history stream, fusion strategy, multi-task prediction heads

• Training practices: scaling, scheduled sampling, regularization (dropout/weight decay), early stopping, and handling missing data

• Uncertainty and calibration: quantile regression heads, ensembles, conformal calibration; producing P10/P50/P90

• Evaluation workflow: rolling-origin backtests; error analysis by hour, regime, and weather conditions

• Practical session: train the hybrid model for multi-horizon forecasts; produce calibrated P10/P50/P90 outputs and compare backtests against baselines

Module 3: Forecast-to-Dispatch MPC and Operational Integration

• Battery modeling for dispatch: SoC bounds, power limits, efficiency, degradation proxy, and reserve requirements

• Optimization design: rolling-horizon MPC driven by probabilistic forecasts; objectives such as arbitrage, ramp smoothing, and peak shaving

• Operating integration: day-ahead vs real-time alignment, penalties, fallback logic for forecast failures, and dispatch fail-safes

• Operationalization: model serving, drift monitoring, shadow deployment, HMI considerations, and KPI tracking

• Practical session: implement MPC dispatch using P50/P10/P90 inputs; generate dispatch/SoC trajectories and a KPI dashboard including cost savings, reserve compliance, curtailment avoided, and VFF

Who Should Enrol

• Data scientists, ML engineers, and MLOps practitioners in energy and utilities

• Power systems and renewables engineers, grid operators, and energy analysts

• Battery/storage planners, asset managers, and virtual power plant (VPP) teams

• Professionals at utilities, IPPs, renewable developers, aggregators, and microgrid operators

• Quant and optimization professionals working on forecasting-to-dispatch workflows

Helpful prerequisites (not mandatory): Python for time-series ML, basics of PyTorch/TensorFlow, familiarity with SCADA and weather datasets, and introductory optimization/MPC concepts.