Peak load forecasting (PLF) is crucial for the power system operation to consider sufficient spinning reserve. Different machine learning algorithms such as Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) are widely used to conduct time-series forecasting. This paper proposes a method based on LSTM integrated with Discrete Wavelet Transform (DWT), which extracts features of power system peak loads, meteorological and calendar data. This paper presents a comparative analysis of the proposed method along with RNN and LSTM. Results show that the DWT-LSTM has a higher R2 and lower error metrics (Mean Squared Error (MSE) and Mean Absolute Error (MAE)) compared to the conventional methods. Comparative studies for DWT-LSTM considering various mother wavelets (Haar, Daubechies, biorthogonal and reverse biorthogonal) are also conducted. Simulation results show that Daubechies wavelet outperforms other mother wavelets. Besides, the filter length and decomposition level of DWT are also essential to conduct the feature extraction.