LocalFarm - Transformer-Based Price Prediction

Using Transformer neural networks to predict local farm produce prices.

Features

• Data Collection: Synthetic data generator / CSV upload / FAO data
• Data Preparation: Automated cleaning, normalization, and sequence creation
• Transformer Model: Multi-head self-attention for time series forecasting
• Evaluation: MAE, RMSE, MAPE, R² with interactive visualizations
• Web Dashboard: Streamlit app with data explorer, trends, training, and predictions

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run the full pipeline (command line)
python main.py

# Run the web dashboard
streamlit run app.py

This opens a browser with 5 pages:

• 📊 Data Explorer — View and analyze the price dataset
• 📈 Price Trends — Interactive historical price charts
• 🤖 Train Model — Configure and train the Transformer with live progress
• 🔮 Predictions — Generate future price forecasts with charts
• 📋 Evaluation — View MAE, RMSE, scatter plots, error distributions

Project Structure

localfarm/
├── data/                    # Price data (CSV)
├── src/
│   ├── __init__.py          # Package init
│   ├── data_collection.py   # Data loading & generation
│   ├── data_preparation.py  # Cleaning, normalization, sequences
│   ├── model.py             # Transformer architecture
│   ├── train.py             # Training loop
│   ├── evaluate.py          # Metrics & visualization
│   └── predict.py           # Inference & forecasting
├── app.py                   # Streamlit web application
├── main.py                  # CLI pipeline runner
├── requirements.txt
└── README.md

Model Architecture

Input (batch, 30, 11 features)
  → Linear Projection → (batch, 30, 64)
  → Positional Encoding
  → 3× Transformer Encoder (4-head attention, d_ff=256)
  → Flatten → FC layers
  → Output (batch, 1)  [next day price]

Input Features (11 total)

Feature	Description
Price	Normalized daily price
day_of_week	Day of week (0-1)
day_of_month	Day of month (0-1)
month	Month of year (0-1)
week_of_year	Week of year (0-1)
price_lag_1	Price 1 day ago
price_lag_7	Price 7 days ago
price_lag_14	Price 14 days ago
rolling_mean_7	7-day rolling average
rolling_std_7	7-day rolling std deviation
rolling_mean_30	30-day rolling average

Evaluation Metrics

Metric	Description
MAE	Mean Absolute Error
RMSE	Root Mean Square Error
MAPE	Mean Absolute Percentage Error
R²	Coefficient of Determination

Technologies

• PyTorch — Deep learning framework
• Streamlit — Web dashboard
• Plotly — Interactive charts
• scikit-learn — Data preprocessing
• pandas/numpy — Data manipulation

How it works

Transformer-Based Time Series Forecasting for Local Farm Produce Price Prediction Step 1 — Data Collection Use datasets such as: -FAO food price data -Kaggle agricultural commodity price datasets -Local market price records (if available) Data format example:

Date Produce Market Price Step 2 — Data Preparation Sort prices by date Handle missing values Normalize price values Convert data into sequences (e.g., last 30 days → next day price)

The above is the requirement for a Final Year University project, How do we implement this project?

Step 3 — Model Development Implement a Transformer-based forecasting model. Train the model on historical price sequences. The model learns relationships between past and future prices.

Step 4 — Evaluation Compare predicted price with real price using: MAE (Mean Absolute Error) RMSE (Root Mean Square Error)

Step 5 — System Development Build a system that: Stores price data Shows price trends Displays predicted future prices Visualizes prediction vs actual price