The energy consumption of air conditioning systems in residential buildings is significantly influenced by the diversity of user behaviors. Traditional statistical features struggle to accurately discern differences in usage patterns under conditions of inadequate data quality, rendering short-term energy consumption prediction models challenged in balancing accuracy and computational cost. Fractal theory, however, offers a distinct advantage in addressing these challenges by quantifying the nonlinear and multi-scale complexity of time series data, thereby enabling the differentiation of user behavior patterns. This study proposes a short-term energy consumption prediction framework that integrates fractal analysis with a data-driven approach. By employing fractal features, users are categorized into low, medium, and high complexity groups, with input data and features optimized across sliding window size, data volume, and feature selection to mitigate noise interference, preserve key patterns, and balance prediction accuracy with computational efficiency.. Based on summer residential air conditioning data, the study finds that sliding window size is negatively correlated with complexity: low-complexity groups prefer longer windows, while medium- and high-complexity groups require shorter ones. For data volume, low- and medium-complexity groups need less, whereas high-complexity groups demand more comprehensive data. In feature selection, low- and medium-complexity groups rely on large-fluctuation features, while high-complexity groups depend on small-fluctuation features. Validation with a 2000-household sample demonstrates a MAPE of 9.82% and a CV-RMSE of 11.40%, achieving a 41.52% error reduction compared to per-household prediction models while significantly saving computational time. By combining fractal features with K-means clustering and LSTM networks, this study overcomes limitations of conventional approaches, offering an efficient solution for short-term energy consumption prediction in residential settings.