air-quality-index-predictions

Implementing a neural network allows customizability for the number of hidden layers and neurons in each layer. Implemented Adam optimizer using both momentum and velocities for weights and biases.

Used OpenWeather Air-Pollution API to get data from Jan 1st, 2015 to Jan 1st, 2025.

Trained the model over the data.

Logistics:

1. From OpenWeather API:
   • Air quality index ranges from 1-5
   • This project follows this encoding:
     • 0 (good): 1-2
     • 1 (moderate): 3-4
     • 2 (bad): 5+
   • This is further encoded using one hot encoding
     • [1, 0, 0] = 0
     • [0, 1, 0] = 1
     • [0, 0, 1] = 2
2. Model accuracy post-training: 92.03%
   • Layout used: Input, HL1, HL2, HL3, Output
   • Neuron count: 8, 20, 15, 10, 3
3. Loss over time