Time Series Forecasting with Deep Learning

Create a deep learning solution for time series forecasting.

Data details:
- Time series type: [UNIVARIATE/MULTIVARIATE]
- Frequency: [HOURLY/DAILY/WEEKLY/etc.]
- History length: [DATA POINTS]
- Forecast horizon: [STEPS AHEAD]
- Seasonality: [PATTERNS PRESENT]

Model requirements:
1. Data preprocessing:
   - Normalization/standardization
   - Sequence creation with sliding window
   - Train/val/test temporal split
2. Architecture options:
   - LSTM with attention
   - GRU networks
   - Temporal Fusion Transformer
   - N-BEATS/N-HiTS
3. Feature engineering:
   - Lag features
   - Rolling statistics
   - Calendar features
4. Training:
   - Teacher forcing
   - Scheduled sampling
   - Multi-step loss
5. Evaluation:
   - MAE, RMSE, MAPE
   - Prediction intervals
   - Backtesting framework
6. Probabilistic forecasting:
   - Quantile regression
   - Monte Carlo dropout

Compare with statistical baselines (ARIMA, Prophet).