Smarter Mobility Challenge

A hierarchical time series forecasting pipeline for electric vehicle charging station occupancy prediction. The system combines gradient-boosted tree models (CatBoost) with deep learning temporal convolutional networks (TCN) via the Darts library, operating at multiple spatial aggregation levels: station, area, and global.

---

Pipeline Overview

The forecasting pipeline executes the following steps in sequence:

1. Baseline forecast — A CatBoost benchmark model is trained and generates initial predictions, stored in Station.pkl.
2. Data ingestion — Raw station-level data is loaded and structured into hierarchical time series via load_data and hierarchy.
3. Covariate construction — Traffic flow data is aggregated from raw CSV files, resampled to 15-minute intervals, scaled, and serialised as covariates.pkl for reuse across runs.
4. TCN model training — A Temporal Convolutional Network (TCN) is fitted independently for each target variable at station level, using cyclic time encodings (hour, day of week, month) as past covariates.
5. Area and global prediction — Forecasts are generated at the area and global aggregation levels.

Output prediction files: Station.pkl, Available.pkl, Charging.pkl, Passive.pkl, Other.pkl.

---

Model Architecture

Baseline: CatBoost (utils/catboost_model.py)

A gradient-boosted tree model trained independently per target variable. Supports both regression (MAE loss) and classification modes. Key hyperparameters:

Parameter	Default
iterations	150
learning_rate	0.1
depth	6
loss_function	MAE

Primary Model: TCN (main.py)

A Temporal Convolutional Network trained on hierarchical time series using the Darts framework. The model is fitted sequentially on two time windows (from L2 then L) to account for temporal distribution shift.

Parameter	Value
Input chunk length	2000
Output chunk length	1921
Kernel size	10
Number of layers	5
Number of filters	10
Dilation base	6
Loss function	L1 (MAE)
Optimiser	Adam (lr = 1e-3)
LR scheduler	StepLR (step=2, γ=0.4)

Cyclic encodings of hour, dayofweek, and month are used as past covariates, scaled via a Darts Scaler.

---

Covariates (utils/covariates.py)

Traffic flow covariates are constructed from hourly CSV files (2020–2021) containing a flow quantity column q. The create_covariates function:

1. Loads and concatenates all raw CSV files from ../0. Input Data/2020/ and ../0. Input Data/2021/.
2. Groups by timestamp and aggregates flow by sum.
3. Resamples to 15-minute frequency with forward-fill for missing values.
4. Serialises the result to covariates.pkl on first run; subsequent calls load from cache.
5. Returns a scaled Darts TimeSeries aligned to the requested start date.