M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

12

channels, grouping by their corresponding transformer ID, and using the Standard Scaler to fit the 255 feature range of the images. With the images in their respective transformer directories, a CNN model was constructed to pass in the images to classify the transformer ID based on the sequential pixels of the meter number image. The CNN model consisted of an Input layer, with a convolutional layer of 32 filters of a 3x3 size and using the ReLU activation function. The input shape is the side of the image length (or 22 for the previously generated meter number images) and a coloured three channel for R, G, and B values. This layer was intended to capture the more basic features in the images and spread the colours seen in the varied power usage per hour. The Max Pooling layer helped to reduce the spatial dimension from the previous layer by half. The additional set of Conv2D and MaxPooling2D layers were intended to capture the additional features within the images. The flattening layer flattened the 3D output of the previous layers into a 1D array for the remaining dense layer. The first dense layer serves as a classification layer using the features extracted by the previous Conv2D and Pooling layers. In contrast, the second dense layer is the output layer that matches the neurons to the number of classes (or transformer IDs) in the dataset. The flattening layer flattened the 3D output of the previous layers into a 1D array for the remaining dense layer. The model was compiled using the ‘Adam’ optimizer, a ‘categorical cross entropy’ loss function, and focused on "Accuracy” for the evaluation metric during the model performance. The CNN model was trained for a set of 300 epochs with a batch size of 32, and the steps per epoch were dynamically calculated based on the sample used divided by the batch size for completeness of training. An additional model was constructed with a dropout layer added, but similar results were achieved. Additional considerations on this approach include altering the model design and dataset to add an LSTM design with the existing CNN model. This approach

103