Curated from public research descriptions, Crossref title/DOI records and official learning resources, checked on 5 November 2026. Paper methods and experimental results still require full-text reading.
01 / Start with the decision
Energy meteorology connects weather information to decisions about power generation and operations. For a first learning task, narrow the scope to a specified site, target and forecast horizon. A point forecast of power, a warning of a rapid change and a maintenance decision require different evidence.
My engineering experience provides a starting point in input validation, asynchronous interfaces and traceable delivery. The new work is to learn the physical meaning of variables, data availability and forecast evaluation before building a larger system.
02 / Foundations: read before adding a model
Read the time-series cross-validation chapter in FPP3 and the ERA5 dataset description. Write down observation time, release time, timezone, units and the distinction between reanalysis and operational forecast products.
ERA5 is useful for retrospective analysis. Downloading it today does not establish that its final values were available at a historical forecast issue time. A retrospective learning exercise and an operational backtest should be labeled separately.
03 / Application papers: three reading questions
- Solar Energy (2024), DOI 10.1016/j.solener.2024.112706: which sources are fused for photovoltaic power prediction, and does each source arrive before the forecast is issued? Record the answer after reading the full text.
- Energy (2024), DOI 10.1016/j.energy.2024.130402: what are the two branches in the wind-power model, and what do single-branch comparisons establish? Check the comparison protocol, not just the best score.
- IEEE GRSL (2023), DOI 10.1109/LGRS.2023.3307717: what is being corrected in the ECMWF forecast products, and how is improvement measured against the uncorrected forecast?
04 / Transfer methods with a baseline
The existing TFT input-availability note is a bridge to multi-horizon forecasting. Treat TFT, feature-level fusion and output correction as choices to compare under the same inputs and evaluation windows. This reading map does not establish that a larger model is needed.
For later reading, keep optimization and AI agents as a separate branch: first identify the action, cost, constraints and validation environment. Forecast quality and policy quality are different questions.
05 / A small, inspectable next step
- Planned: make a data dictionary with source licenses, units, timestamps and missing-value rules.
- Planned: compare persistence and same-time-previous-day baselines on chronological windows; retain errors by horizon and the sample count.
- Planned: show one successful result and one delayed-input case in a small dashboard. No new dataset or model reproduction is claimed in this note.
Sources & further reading
- Application of multi-source data fusion on intelligent prediction of photovoltaic power · Solar Energy, 2024
- Research on the short-term wind power prediction with dual branch multi-source fusion strategy · Energy, 2024
- A Deep Learning Method Integrating Multisource Data for ECMWF Forecasting Products Correction · IEEE GRSL, 2023
- Forecasting: Principles and Practice · Time series cross-validation
- Copernicus CDS · ERA5 hourly data on single levels
Compiled from code and documents, with AI-assisted drafting and translation. Examples and proposed checks are learning material; only linked evidence establishes what was run.