Design end-to-end regression for energy demand
Overview
This question evaluates a candidate's ability to design and justify an end-to-end regression system for next-day site-level energy demand forecasting, testing competencies in time-series modeling, feature engineering, regularization, evaluation metrics and SLAs, production ML pipelines, monitoring, explainability, and data privacy.
Amazon
Oct 13, 2025, 9:49 PM
Data Scientist
Onsite
Machine Learning
0
0
End-to-End Daily Energy Prediction for Commercial Buildings
Context
You are asked to design and justify an end-to-end regression system that predicts next-day daily site-level electricity consumption (kWh) for a portfolio of commercial buildings across multiple years (2019–2025). The system should support forecasting for each building ("site") using:
- Hourly smart-meter reads
- Weather: temperature, humidity, wind
- Calendar and holiday flags
- Building metadata: floor area, vintage (year built), HVAC type
- Optional external features: day-ahead electricity price, outage alerts
Assume you must deliver both accurate forecasts and a robust production pipeline suitable for enterprise operations.
Requirements
- Formulate the problem and propose baselines.
- Engineer features for seasonality, interactions (e.g., temperature × HVAC), and occupancy proxies.
- Choose and justify regularization; address multicollinearity; detect and mitigate heteroscedasticity.
- Use time-series-aware cross-validation and avoid leakage; be explicit about any lag/rolling constructs.
- Specify metrics (e.g., RMSE, MAPE) and business-facing SLAs (e.g., billing tolerance bands).
- Handle missing/corrupted sensors and concept drift across years (2019–2025).
- Productionize: outline training and inference pipelines, model versioning, and monitoring (data/feature drift, residual shifts, retraining triggers).
- Explainability for non-ML stakeholders (global vs. local) and safe failure modes.
- Security and privacy constraints for tenant data.
- Propose an ablation plan to quantify incremental value of external features and describe a backtest on 2023–2024 with 2025 held out for final evaluation.
Solution
Login required
Show