This is my first ever ML project, where I had to learn about CIGRE-MV Networks, Pandapower implementation of it, DLPF Solver, DNN implementation of Non-Linear Newton Raphson equation, 2 Ygdrassil Random Forest Classifiers and a comparator component to compare between DLPF and DNN.
- Underfitting means the model is too simple for the complex dataset, hence performing poor on both train and test datasets!
- Overfitting means the model is too flexible for the simple dataset, which leads it to learn the pattern of noises instead of actual data and memorising it.
- Underfitting => Model Capacity < Data Complexity
- Overfitting => Model Capacity>> Data Complexity (Noise Included)
- Bias is the error that occurs due to simplification of the function
- Variance is the error that stems from the high sensitivity of model to noise and other changes.
| Layer Name | Type | Output Shape | Parameters |
|---|---|---|---|
| dense_hidden_1 | Dense | (None, 512) | 33,280 |
| bn_1 | BatchNormalization | (None, 512) | 2,048 |
| relu_1 | Activation (ReLU) | (None, 512) | 0 |
| dense_hidden_2 | Dense | (None, 256) | 131,328 |
| bn_2 | BatchNormalization | (None, 256) | 1,024 |
| relu_2 | Activation (ReLU) | (None, 256) | 0 |
| dense_hidden_3 | Dense | (None, 128) | 32,896 |
| bn_3 | BatchNormalization | (None, 128) | 512 |
| relu_3 | Activation (ReLU) | (None, 128) | 0 |
| output_layer | Dense | (None, 88) | 11,352 |
- Total Parameters: 212,440 (~829.84 KB)
- Trainable Parameters: 210,648 (~822.84 KB)
- Non-trainable Parameters: 1,792 (~7.00 KB)
- 3 fully connected hidden layers (512 → 256 → 128)
- Each hidden layer followed by:
- Batch Normalization
- ReLU activation
- Final output layer:
- 88 neurons (Multi-Class Classification)
- Batch Normalization improves training stability and convergence.
- ReLU introduces non-linearity.
- Model size is moderate (~212K params).