Note: This repository was revised from an internal project at a previous company (CADS - FPT) and cleaned for portfolio sharing. The demo uses synthetic data for illustration only; real electricity consumption data is not included.
- 13 provinces across Central Vietnam — province-level monthly electricity consumption forecasting
- < 6% MAPE across all 13 provinces; < 4% MAPE for 6 of 13 — evaluated over 18 months
- Decomposition-based approach — trend (linear), seasonality (statistical model), and residual (ARIMA) modeled independently then combined
- Per-month parameter tuning — each calendar month uses an independently optimized parameter set (lag, order1, order2, shift)
- Dual forecast horizons — next-month (T+1) and recursive 12-month (T+12) predictions
- Breakdown forecasting — sub-series analysis by district and industry sector validated against total-consumption baseline
A self-contained demo (demo_prediction.py) illustrates the full pipeline using synthetic data:
python3 demo_prediction.pyThe demo generates component-level forecasts and saves:
| Output | Description |
|---|---|
docs/demo_forecast_next_month.csv |
T+1 forecast with trend, seasonal, residual, and final prediction |
docs/demo_forecast_12_months.csv |
Recursive T+12 forecast with component breakdown |
docs/forecast_sample.png |
Combined plot — history + predictions for example provinces |
docs/forecast_<province>.png |
Individual province-level forecast plots |
The model separates each province's consumption into three interpretable components:
Total forecast vs. historical:
Component breakdown (trend + seasonal + residual):
Evaluation period: January 2022 – July 2023 (18 months) · 13 provinces · Data from 2014–2023
| Metric | Result |
|---|---|
| Primary metric | MAPE (Mean Absolute Percentage Error) |
| All 13 provinces | Average MAPE < 6% |
| 6 of 13 provinces | Average MAPE < 4% |
| 3 provinces | Near the 6% threshold (DANANG 5.8%, DNONG 5.4%, QNGAI 5.5%) |
Provinces fell into two groups based on error stability:
| Group | Provinces | Characteristics | Action |
|---|---|---|---|
| Group 1 | BINHDINH, GIALAI, KHANHHOA, KOMTUM, PHUYEN, QUANGNAM, QUANGBINH, QUANGTRI, TTHUE | Stable, low error (1.7–4.6% mean absolute error) | Monitor |
| Group 2 | DANANG, DAKLAK, DAKNONG, QUANGNGAI | Higher variance, less stable (5.1–6.4% mean) | Active treatment |
Beyond total-consumption forecasting, the model was evaluated on decomposed sub-series:
By Industry Sector (NN_LV1):
| Sector | Performance |
|---|---|
| Residential (Sinh hoạt dân dụng) | Best — 2.5–8.9% MAPE, stable |
| Other activities | Good — low and stable error |
| Industry & Construction | Moderate — mostly <10% (except DLAK, DNONG) |
| Agriculture, Forestry, Fishery | High — 15–20% MAPE, volatile (esp. Central Highlands) |
| Commerce, Hotels, Restaurants | Highest — 15–20%+, most volatile across all provinces |
By District (DVDC_LV3):
- ~65% of district-level sub-series achieved < 10% MAPE
- Total-consumption forecasting vs. sum-of-sub-series forecasting produced comparable results — validating the decomposition approach
- District-level breakdown was selected as the preferred direction for deployment due to more uniform error distribution
Evaluation period: December 2022 – July 2023 (8 months)
- T+1 error: ~6% → T+12 error: ~14%
- Error increases gradually with horizon distance — expected for recursive strategy
- Four strategies evaluated: Recursive, Multi-output, Direct, DirRec — Recursive chosen for simplicity and acceptable error propagation
Goal: Forecast monthly electricity consumption at the province level for 13 Central Vietnam provinces using historical data from 2014 onward.
Client: CPC (Central Power Corporation) — a major regional electricity utility
Each province's monthly consumption series is decomposed into three independent components, forecast separately, then recombined:
Final Prediction = Trend Forecast + Seasonal Forecast + Residual Forecast
| Component | What it captures | Model |
|---|---|---|
| Trend | Long-term growth, decline, or stability | Linear model |
| Seasonal | Repeated monthly and yearly cycles | Statistical model with per-month parameters |
| Residual | Short-term variation unexplained by trend or seasonality | ARIMA |
Captures the underlying direction of each province's consumption. Simple and robust — avoids overfitting to short-term noise.
The most heavily engineered component. Uses historical consumption and cumulative consumption to estimate the next month's value. A key insight from the project: cumulative consumption patterns are highly correlated across years (correlation ≈ 1), making cumulative forecasting a reliable anchor.
Parameters tuned independently for each calendar month (January–December):
| Parameter | Role |
|---|---|
lag |
Number of historical months used for estimation (range: 3–12) |
order1 |
Function order for monthly consumption estimation (max: 5) |
order2 |
Function order for cumulative consumption estimation |
shift |
Steps to shift historical data backward (pattern matching) |
Parameter selection: grid search over parameter space, selecting the combination with lowest MAPE on the training set (2014–2021). A critical finding: identifying which past year's pattern the forecast year most resembles is more important than parameter tuning — correlation-based year matching was used for provinces like DANANG.
Captures remaining short-term patterns after trend and seasonality are removed. Added in Version 4 of the pipeline — prior versions assumed residual = 0.
For multi-month horizons, the model chains one-step predictions:
Predict T+1 → feed into input → predict T+2 → ... → predict T+12
Four strategies were evaluated (Recursive, Multi-output, Direct, DirRec). Recursive was selected — error propagation is acceptable (~6% → ~14% over 12 steps) and only one model needs to be maintained.
# Install dependencies
pip install -r requirements.txt
# Run the demo (synthetic data, no setup needed)
python3 demo_prediction.py# T+1 forecast (next month)
python main/forecast_t1_13pr_sum.py \
-f 01/2023 -t 03/2023 \
-p /path/to/params \
-i /path/to/input.parquet \
-o /path/to/output_t1.parquet
# T+12 forecast (next 12 months)
python main/forecast_t12_13pr_sum.py \
-f 01/2023 -t 03/2023 \
-p /path/to/params \
-i /path/to/input.parquet \
-o /path/to/output_t12.parquet| Argument | Description |
|---|---|
-f |
Start month (MM/YYYY) |
-t |
End month (MM/YYYY) |
-p |
Parameter file for 13 provinces |
-i |
Input electricity data (CSV or Parquet) |
-o |
Output path for forecast results |
| Column | Description |
|---|---|
Date |
Observation month (YYYY-MM-DD recommended) |
Province |
Province name |
Consumption |
Electricity consumption value |
T+1 results include Province, Year, Month, y_tr (actual), and y_pr (predicted). T+12 results include columns y_pr1 through y_pr12 for each forecast horizon.
.
├── demo_prediction.py # Self-contained demo (synthetic data)
├── demo_forecast_next_month.csv # T+1 demo output
├── demo_forecast_12_months.csv # T+12 demo output
├── requirements.txt
├── docs/ # Generated forecast plots
├── jupyter-notebook/
│ └── CPC_Forecast.ipynb # Interactive forecasting notebook
└── main/
├── forecast_t1_13pr_sum.py # T+1 forecasting script
├── forecast_t12_13pr_sum.py # T+12 recursive forecasting script
└── helper.py # Shared utilities
- Update input data paths before each forecast run.
- T+12 uses recursive forecasting — errors compound over longer horizons (~6% at T+1 → ~14% at T+12).
- The statistical model depends heavily on matching the forecast year to a historically similar year; correlation-based year selection is critical for accuracy.
- Cross-province data pooling was tested but did not improve results — each province has distinct consumption patterns.
- This repository was revised from original internal project code; internal paths and data have been removed.