Effective energy consumption forecasting in smart homes is vital for optimising resource usage and integrating renewable energy sources. Current research is often hindered by factors such as limited or low quality data, the challenge of identifying suitable prediction models, the variability of consumption patterns, and the scalability limitations of these models. In this study, we examine the scalability and performance of three advanced machine learning frameworks: Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Artificial Neural Networks (ANNs) for short-term energy consumption forecasting across multiple dataset sizes. Specifically, it used four generated datasets representing 20, 50, 100, and 200 smart homes, each covering 365 days of energy consumption data. We assess how well each model adapts to growing dataset sizes by measuring root mean square error (RMSE), mean absolute error (MAE), coefficient of determination ( \({R}^2\) ), training time, and inference speed. We examine the models’ ability to generalize from smaller to larger datasets and their suitability to capture diverse consumption patterns in different household datasets. In addition, we assess the resource and time efficiency of each model. Our findings show that ANN models deliver reliable and precise predictions for energy consumption, making them particularly suitable for both residential-scale and city-wide smart energy management.