An ADMM-LSTM framework for short-term load forecasting.

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Accurate short-term load forecasting (STLF) is essential for power system reliability and efficiency.
• Increasing data complexity from renewables and EVs necessitates advanced forecasting methods.
• Traditional long short-term memory (LSTM) training methods face limitations like exploding/vanishing gradients.

Purpose of the Study:

• To present an innovative LSTM optimization framework, ADMM-LSTM, for STLF.
• To address the limitations of conventional stochastic gradient-based training methods for LSTMs.
• To enhance the accuracy and computational efficiency of STLF.

Main Methods:

• Developed ADMM-LSTM, a distributed training framework for LSTMs using the alternating direction method of multipliers (ADMM).
• Introduced a novel backward-forward parameter update order to reduce computational time.
• Utilized proximal point algorithm or local linear approximation for subproblem solutions, avoiding external solvers.
• Provided theoretical analysis of ADMM-LSTM's convergence properties.

Main Results:

• ADMM-LSTM effectively trains LSTM networks in a distributed manner.
• The backward-forward update order significantly reduces computational time.
• The framework inherently avoids exploding or vanishing gradient issues.
• Experimental results on public datasets show superior performance compared to existing STLF methods.

Conclusions:

• ADMM-LSTM offers a robust and efficient alternative for STLF.
• The proposed optimization framework enhances LSTM performance in power system applications.
• The gradient-free nature and convergence properties make ADMM-LSTM a valuable tool for complex forecasting tasks.