Electricity load forecasting is essential for economic viability and reliable operations in modern, complex power systems. Although traditional statistical methods and deep learning methods can be utilized for predictions, they tend to fall short in capturing complex non-linear behaviors with increasing demand. For this reason, in this study, a new cascaded method, Multi-Stage Deep Learning Framework (MSDL) is proposed which can capture these behaviors more accurately. MSDL consists of three different architectures, namely Seasonal Decomposition (SD), Empirical Mode Decomposition (EMD), and Long Short-Term Memory (LSTM). In MSDL, the time series data is first decomposed into trend, seasonal, and residual components using SD. The residual component is then further decomposed into Intrinsic Mode Functions (IMFs) using EMD to extract intricate, high-frequency patterns. These components are subsequently used as inputs to LSTM models for forecasting. Benchmarking results demonstrate that the proposed method outperforms several reference models in terms of RMSE, MAE, and MAPE. The study also goes beyond performance comparison by investigating the influence of the quantity of IMFs on forecast accuracy. It has been shown that for the selected case study, including more than five IMFs in forecasting does not significantly increase prediction performance.