AI-Based Short-Term Load Forecasting Using CNN-LSTM

n this project, I led a group of 3 students to create a deep learning model that predicts 7-days ahead electric load in Nineveh, Iraq. We used historical electrical load data of the Nineveh power grid and weather data from Open-meteo API to improve prediction accuracy. The project uses a combination of CNN and LSTM layers.

This work was developed as part of our B.Sc. graduation thesis.

Model Architecture

Input → Conv1D (336 filters, kernel=4, tanh) → MaxPooling1D → Dropout(0.2)

→ LSTM (48 units) → Dropout(0.3) → Flatten → Dense(1)

Input Features

The model uses a 7-day lookback window with the following features:

Load Features: Average load, supply hours, cut hours

Weather Features: Temperature (mean, max, min), humidity, precipitation, wind speed, pressure, cloud cover

Calendar Features: Day of month, month, day-type encoding