Transformer-LSTM Hybrid Forecast Engine for RE + Storage Dispatch

The Transformer–LSTM Hybrid Forecast Engine for RE + Storage Dispatch workshop is a three-day, hands-on sprint from data prep to operations. You’ll build time-aligned SCADA+weather datasets, train a hybrid Transformer (weather) + LSTM (plant history) model with calibrated multi-horizon outputs (P10/P50/P90), and evaluate via rolling backtests. Finally, you’ll drive battery dispatch using rolling-horizon MPC, producing dispatch/SoC trajectories and a KPI dashboard (cost savings, reserve compliance, curtailment avoided, VFF).

Enable participants to build and deploy a Transformer–LSTM hybrid that delivers calibrated multi-horizon RE forecasts (P10/P50/P90) and converts them into optimal battery dispatch via rolling-horizon MPC, with measurable KPI gains.

• Build time-aligned SCADA+weather datasets with time-aware splits

• Engineer features; establish persistence/LSTM baselines

• Train a late-fusion Transformer–LSTM with P10/P50/P90 outputs

• Ensure training hygiene and uncertainty calibration

• Evaluate via rolling backtests; create a compact model card

• Convert forecasts to MPC-based battery dispatch under constraints

• Operationalize (serving, drift, shadow runs, HMI) and track KPIs

📅 Day 1 – Data & Hybrid Basics

• Signals & horizons: solar/wind, net load, price; intraday/day-ahead use cases
• Time-aware splits; feature engineering (lags/rolls, weather look-ahead, plant metadata, calendar)
• Metrics: RMSE/sMAPE, multi-horizon, pinball loss; baseline design (persistence)
• Hybrid idea: Transformer (exogenous weather) + LSTM (plant history) with late fusion
• Hands-on: Build a time-aligned SCADA+weather table and train persistence + single-layer LSTM baselines

📅 Day 2 – Train the Transformer–LSTM & Calibrate

• Architecture: seq lengths, encoder–decoder attention, LSTM history stream, fusion, multi-task heads
• Training hygiene: scaling, scheduled sampling, dropout/WD, gap handling, early stop
• Uncertainty: quantile heads, ensembles, conformal calibration; P10/P50/P90 outputs
• Evaluation: rolling-origin backtests; error by regime/hour
• Hands-on: Train the hybrid model for multi-horizon forecasts; emit P10/P50/P90 and a compact backtest vs baselines

📅 Day 3 – Forecast→Storage Dispatch (MPC) & Ops

• Battery model: SoC bounds, power limits, efficiency, degradation proxy, reserves
• Optimization: rolling-horizon MPC using forecast ensembles; objectives—arbitrage, ramp smoothing, peak shaving
• Operations: DA/RT alignment, penalties, fail-safes for bad forecasts; MLOps (serving, drift, shadow runs, HMI)
• KPIs: cost savings, reserve compliance, curtailment avoided, VFF (value of forecast)
• Hands-on: Implement MPC dispatch consuming P50/P10/P90; output dispatch/SoC trajectories and KPI dashboard

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

• Power system/renewables engineers, grid operators, and energy analysts

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

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

• Quant/optimization folks working on forecasting→dispatch workflows (DA/RT)

• Prereqs (helpful, not mandatory): Python + time-series ML; basics of PyTorch/TF; familiarity with SCADA/weather data; introductory optimization/MPC concepts

• Clean, time-aligned SCADA+weather dataset with strong features

• Calibrated Transformer–LSTM multi-horizon forecasts (P10/P50/P90)

• Robust evaluation: rolling backtests, regime/hour diagnostics, model card

• Forecast uncertainty managed (quantiles, ensembles, conformal)

• Battery dispatch via rolling-horizon MPC with SoC/limits/efficiency

• End-to-end pipeline to KPIs: cost savings, reserve compliance, curtailment avoided, VFF