☀️ Suncast — Hourly Solar PV Generation Forecasting Model (China Region)

A machine learning model that predicts hourly solar PV power generation (kWh) for any location across mainland China, given latitude, longitude, and a date range.

📌 Model Overview

Item	Detail
Task	Tabular Regression (Solar Irradiance → PV Power)
Algorithm	Random Forest Regressor (via PyCaret AutoML)
Target Region	Mainland China (UTC+8)
Temporal Resolution	1-hour intervals
Output Unit	kWh (1 kW standard PV plant)
Training Period	2024 full year
Training Samples	4,861,296

📊 Performance

Metric	Value
MAE	76.19 W/m²
RMSE	126.96 W/m²
R²	0.748
MAPE	1.49%

Notable observations:

✅ High accuracy during summer months (abundant solar irradiance)
⚠️ Increased error in winter (low irradiance, high meteorological variability)
The seasonal structure of the model allows for long-term extensibility

🗂️ Data Sources

Input — GFS (Global Forecast System, NOAA)

Spatial resolution: 1° × 1°
Temporal resolution: 1 hour
Coverage: Lat 19°–53° (2° step), Lon 74°–134° (2° step) → 558 grid points

Variable	Unit
Surface Pressure	Pa
Surface Temperature	K
Relative Humidity (2m)	%
U-Component of Wind (10m)	m/s
V-Component of Wind (10m)	m/s
Sunshine Duration	s
Low / Mid / High Cloud Cover	%
Downward Short-Wave Radiation Flux	W/m²

GFS DSWRF is a model-simulated value computed via the RRTMG radiation transfer scheme — not a direct satellite measurement.

Target — NASA POWER / CERES SYN1deg

Source: CERES SYN1deg (Ed4.x), cross-calibrated with Terra/Aqua CERES, MODIS, and GEO satellites
Spatial resolution: 1° × 1° (downsampled to 2° × 2°)
Temporal resolution: 1 hour (linearly interpolated from 3-hour data)
Time zone: UTC+8 fixed (unified across all of China)

🧠 Model Training Details

Feature Engineering

Spatiotemporal alignment and standardization of GFS input variables
Added temporal features: hour_local, month_local, day_of_year, season

Candidate Models Compared

Extra Trees Regressor
Random Forest Regressor ✅ (selected)
LightGBM
Gradient Boosting Regressor

Random Forest was selected for its strong resistance to overfitting and balanced performance across all evaluation metrics.

Training Configuration

Setting	Value
Train / Test Split	80% / 20%
Cross-Validation	k-fold (k=10)
Hyperparameter Tuning	Grid Search

⚡ PV Power Conversion

Predicted solar irradiance (W/m²) is converted to power generation (kWh) using pvlib.

Parameter	Value
Panel Tilt	25°
Panel Azimuth	180° (south-facing)
Temperature Coefficient	−0.004 /°C
Capacity	1 kW (standard)

Power generation is set to 0 kWh before 06:00 and after 19:00 (local time).

🚀 How to Use

from pycaret.regression import load_model, predict_model
import pandas as pd

# Load model
model = load_model("AutoML_model_v1")

# Prepare input features
input_data = pd.DataFrame([{
    "sp": 101325,       # Surface Pressure [Pa]
    "t": 300.15,        # Surface Temperature [K]
    "r2": 60.0,         # Relative Humidity [%]
    "u10": 2.0,         # U-Wind [m/s]
    "v10": -1.5,        # V-Wind [m/s]
    "SUNSD": 3200,      # Sunshine Duration [s]
    "lcc": 10.0,        # Low Cloud Cover [%]
    "mcc": 5.0,         # Mid Cloud Cover [%]
    "hcc": 20.0,        # High Cloud Cover [%]
    "sdswrf": 650.0,    # DSWRF [W/m²]
    "hour_local": 12,
    "month_local": 7,
    "day_of_year": 190
}])

# Predict irradiance
prediction = predict_model(model, data=input_data)
print(prediction["prediction_label"])

📁 Repository Files

File	Description
`Suncast_v1.pkl`	Trained PyCaret Random Forest pipeline

⚠️ Limitations

Training data is limited to 2024 only (originally planned for 2020–2024; reduced due to GFS server instability and storage constraints)
Grid resolution is 2° × 2° — predictions use the nearest grid point to the input coordinates
Not applicable outside mainland China grid coverage

📄 License

This model is released under the Apache 2.0 License.

Downloads last month: -; Downloads are not tracked for this model. How to track