Vision-Based Smart Meter Reading and Energy Consumption Analytics Using AI

Abstract

In this work, a deep learning-based energy consumption predictor with Long Short-Term Memory (LSTM) and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) network models are applied to time-series data. This is aimed at enhancing the accuracy of prediction by ensuring that both short-term variation and the temporal dependencies of electricity demand are in the model. Kaggle provides public datasets such as hourly energy consumption data, which are analyzed. The cleaning, chronological sorting, Min-Max scaling, and sliding window sequence generation (24-time steps) preprocess the data. To extract temporal patterns, including trends and seasonality, feature engineering and exploratory data analysis are used. Training of models is done with the help of 80: 20 time based split to avoid the data leakages and it is also evaluated with the help of Mean Squared error (MSE), root mean squared error (RMSE) and loss curves. Experimental ablation reveals that LSTM has an ultimate training loss of approximately 0.000175 and validation loss of approximately 0.000060 whereas CNN-LSTM has 0.000176 and 0.000058 accordingly. CNN-LSTM model has a little better performance because it has an enhanced feature extraction and noise reduction. Both models demonstrate stable convergence, low overfitting, and good generalization. Additionally, 24 hours recursive forecasting creates smooth and realistic forecasts that are consistent with the historical trends. On the whole, the suggested CNN-LSTM method is a more appropriate solution to short-term energy consumption forecasting due to its higher accuracy and strength, which is why it can be used in smart grid and energy management.