The advanced metering infrastructure (AMI) loads are affected by changes in the temperature, humidity, and energy consumption of electrical equipment. Due to the high variability of AMI loads, the operating risk of a power grid caused by prediction error must be addressed. This paper utilizes discrete wavelet transform (DWT) to decompose load signals into low- and high-frequency components. Unnecessary high-frequency signals are eliminated by appropriately reconstructing signals that increase the accuracy of the forecasting model. Signal reconstruction is a combinatorial optimization process. This paper further integrates a grey wolf optimizer (GWO) and an autoregressive with multiple exogenous inputs (MIARX) model to find the optimal solution for signal reconstruction. When the day-ahead hourly forecasting of AMI loads is obtained, a quantile regression (QR) model is utilized to produce asymmetric prediction intervals. An index that considers both prediction interval coverage probability (PICP) and an evaluation resolution of criterion (ERC) is used to evaluate the performance of the obtained prediction intervals. To verify its feasibility, the proposed method is tested on smart AMI users in a green energy building located at Cheng Kung University in Taiwan.