<![CDATA[Automatic interpretation of building electrical metric data is essential for assessing the reliability and suitability of electrical systems. The emergence of Large Language Models (LLMs) has created new opportunities to automate the inspection and deterministic interpretation of these metric values without requiring manual input. This study evaluates the performance of local computing (Edge AI) in interpreting and classifying electrical system status using a Zero-Shot Learning approach without the need for model retraining. The interpretation rules were based on the PUIL 2020 standard and included parameters such as voltage deviation, frequency, current load, imbalance, and power factor. The comparative evaluation involved two 8-bit quantized models: Llama 3.1 (8B) and Qwen 2.5 (7B), tested using 200 historical building electrical panel data samples (100 normal and 100 anomalous). The assessment covered LLM performance metrics (syntactic and semantic accuracy), anomaly detection classification, and hardware resource efficiency. The results show that Qwen 2.5 (7B) outperformed Llama 3.1 (8B) in mathematical reasoning tasks, achieving an accuracy of 91.50% and a precision of 95.60%, with minimal false positives. In addition, Qwen completed the analysis 42 minutes faster while using a peak RAM consumption of 8.9 GB. In contrast, Llama 3.1 demonstrated excessive sensitivity, resulting in an accuracy of 57.50%, a precision of 54.19%, and higher memory usage (11.9 GB). These findings indicate that the effectiveness of Zero-Shot Learning in LLMs for logical reasoning tasks depends more on the model’s training bias than on the number of parameters. Models specifically trained for programming and mathematical reasoning, such as Qwen 2.5, are more reliable, consistent, and efficient in interpreting electrical metrics compared to general conversational models.]]>
