Real-time anomaly detection in energy consumption is crucial for identifying technical inefficiencies and user behavior issues that lead to energy waste. Traditional methods rely on utilizing large historical consumption patterns, but data limitations in certain domains hinder the application of these systems. This study addresses this challenge by leveraging transfer learning with long short-term memory networks, using school building energy datasets as a source domain to improve performance in data-scarce target domains. The evaluated models were trained to forecast 8-h energy consumption and detect anomalies. Results show that transfer learning models trained with 40 % of the dataset generalize better, reducing sensitivity to minor fluctuations and lowering false alarm rates compared to baseline models which trained on full training dataset. Those models tend to overfit to small variations, which led to increased false positives. These findings highlight the transfer learning effectiveness in improving anomaly detection reliability, ensuring models focus on consistent and persistent changes in consumption patterns.