Business Context

VoltSense operates 18,000 industrial IoT sensors across manufacturing plants and uses hourly forecasts to trigger maintenance alerts. The current forecasting pipeline is unstable because raw sensor streams contain spikes, dropouts, and short-term measurement noise.

Dataset

You are given a historical univariate-to-multivariate time series dataset for power_load_forecasting. Each row is one sensor-hour observation, and the goal is to predict the next 24 hours of power_load_kw while making the training data robust to noise.

• Size: 2.4M hourly rows from 18,000 sensors over 18 months, 40 features
• Target: Regression — power_load_kw for the next 24 hours
• Noise characteristics: ~6% spike outliers, 9% missing intervals, occasional level shifts after sensor recalibration
• Missing data: 9% missing in raw sensor channels, concentrated around network outages

Success Criteria

A good solution should reduce forecast error relative to a raw-data baseline and remain stable during noisy periods. Target MAE < 2.8 kW, RMSE < 4.1 kW, and at least 15% lower MAE on noisy windows versus a model trained on unsmoothed inputs only.

Constraints

• Batch inference must score all sensors hourly in under 10 minutes
• The operations team needs interpretable preprocessing steps
• Retraining budget supports weekly retraining, not daily

Deliverables

1. Propose techniques to detect and handle noisy observations in the time series
2. Build a regression pipeline with explicit denoising and feature engineering
3. Use a time-aware validation strategy that avoids leakage
4. Compare at least one baseline against a denoised approach using business-relevant metrics
5. Explain how the approach would be deployed and monitored in production