Weather-Based Power Load Forecasting

A deep learning project that forecasts hourly power load using a multi-region LSTM (Long Short-Term Memory) neural network. The model leverages energy consumption data from multiple PJM (Pennsylvania-New Jersey-Maryland) regions to predict power demand for a target region.

Overview

This project implements a time series forecasting pipeline that:

• Ingests hourly energy consumption data from 13+ regional datasets (AEP, COMED, DAYTON, DEOK, DOM, DUQ, EKPC, FE, NI, PJME, PJMW, PJM_Load, pjm_hourly_est)
• Preprocesses and merges multi-region data with timestamp alignment and interpolation
• Trains a 2-layer LSTM model with 128 hidden units to predict the next hour's load
• Evaluates forecasts using RMSE, MAE, and MAPE metrics

The target region for forecasting is PJME (PJM East), with other regions used as auxiliary features to capture cross-regional load patterns.

Project Structure

Weather-Based Power Load Forecasting/
├── weather-based-power-load-forecasting.ipynb   # Main Jupyter notebook
├── systemdiagram.drawio.png                     # Pipeline architecture diagram
├── output.png                                   # Sample forecast visualization
├── Weather-Based Power Load Forecasting Report.docx
├── requirements.txt                             # Python dependencies
├── README.md                                    # This file
└── venv/                                        # Virtual environment

System Architecture

The pipeline follows four main stages:

1. Data Preparation — Raw CSV files → Merge & align timestamps → Interpolate missing values → MinMax scaling (0–1)
2. Sequence Creation — Create input sequences (24-hour lookback) and targets (1-hour ahead)
3. LSTM Architecture — 2 LSTM layers (128 units each), dropout (0.2), linear output layer
4. Post-Processing & Evaluation — Inverse transform predictions → Compare forecast vs actual

Data Source

The project uses the Hourly Energy Consumption dataset from Kaggle, which contains hourly power load (MW) for multiple PJM regions.

Data files (place in ./data/ or Kaggle input path):

• AEP_hourly.csv, COMED_hourly.csv, DAYTON_hourly.csv, DEOK_hourly.csv
• DOM_hourly.csv, DUQ_hourly.csv, EKPC_hourly.csv, FE_hourly.csv
• NI_hourly.csv, PJME_hourly.csv, PJMW_hourly.csv
• PJM_Load_hourly.csv, pjm_hourly_est.csv

Requirements

• Python 3.11+
• PyTorch (CPU or CUDA)
• pandas, numpy, scikit-learn, matplotlib

Setup

1. Clone or download this repository.

2. Create a virtual environment (recommended):

   python -m venv venv
   venv\Scripts\activate   # Windows
   # source venv/bin/activate   # Linux/macOS

3. Install dependencies:

   pip install -r requirements.txt

4. Obtain the dataset from Kaggle and either:

   • Place CSV files in a data/ folder and update paths in the notebook, or
   • Run the notebook on Kaggle (paths are pre-configured for Kaggle input).

Usage

1. Open weather-based-power-load-forecasting.ipynb in Jupyter Notebook or JupyterLab.
2. Ensure the data paths in the Config class point to your dataset location.
3. Run all cells to load data, train the model, and view results.

Configuration

Key parameters in the Config class:

Parameter	Default	Description
SEQ_LENGTH	24	Lookback window (hours)
PREDICT_HORIZON	1	Steps ahead to predict
HIDDEN_SIZE	128	LSTM hidden units
NUM_LAYERS	2	Number of LSTM layers
EPOCHS	20	Training epochs
TARGET_REGION	PJME	Region to forecast

Model

• Architecture: MultiRegionLSTM — LSTM with configurable input size, hidden size, and layers
• Loss: MSE (Mean Squared Error)
• Optimizer: Adam
• Metrics: RMSE (MW), MAE (MW), MAPE (%)

Output

The notebook produces:

• Training loss per epoch
• Final RMSE, MAE, and MAPE on the test set
• A plot comparing actual vs predicted load for the first 500 test hours