You already know: time-series forecasting and quantile losses (D3), gradient-boosted trees (D3), leakage and point-in-time joins (D16), and from the surge case that forecasts drive decisions with deadlines. This case adds forecasting for physical systems: the inputs (weather) are themselves forecasts, errors cost money in a market, the hardest days are the ones outside your training data, and a forecast must be submitted on time whatever happens. The design questions: which horizons and levels, how to use weather forecasts honestly, which models for which horizon, and how to make forecasts dependable in operations.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "train an LSTM on historical load and temperature". A staff candidate first asks which horizons and deadlines matter, at which level of the grid, how errors are paid for, and what drives load. Those answers produce one model family per horizon with a submission deadline, hierarchical forecasts that add up, quantile forecasts for reserves, and weather forecasts as they were issued, never observed weather, in training and testing.
Assumptions you state: