Deep Learning and Time Series Forecasting for Smarter Energy
summary
A production forecasting pipeline built on pandas scikit tensorflow ingests ~77,000 smart meters to train shared neural models with household embeddings, mixture‑density outputs and a copula layer for tail correlations. It outputs probabilistic one‑step and iterated portfolio scenarios, feeds trading and battery optimization, handles negative price events, and runs as a dockerized fit/predict model with frozen fine‑tuning of embeddings. Results show clearer aggregate forecasts though individual spikes remain hard to predict.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:59Energy grid challenges
- 6:53Time series forecasting
- 11:30Deep learning for forecasting
- 13:57Feature engineering
- 16:23Building neural networks
- 19:17Model evaluation
- 21:20Forecast errors handling
- 25:52Probabilistic forecasting
- 28:21Scenario sampling
- 30:28Tail risk modeling
- 31:51Model implementation
- 32:12Key takeaways
- 35:28Q&A
related talks
29mA dive into time series for the energy sector
37mEstimating solar production w/ distributed deep learning from smart meter data
34mTime series for scikit-learn people
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
01h 02mDeep learning for time series forecasting and classification in practice
28mEfficient Time-Series Forecasting with Thousands of Local Models on Databricks
23mDataEngBytes 2023 - MEL-T3-01 - Rajath Vanikul
45mForecasting at Scale