Abstract

Abstract
Quantitative Trading in North American Power Markets
Dominic Isaac Andoh
Short-term load forecasting (STLF) in electrical grids is critical for efficiency and reliability. Many
countries in the west have deregulated their electric power industry, allowing for a free and competitive
market; this has made load forecasting a more critical task for estimating future spot prices. Load
forecasting is a complex task due to seasonal variation and the non-stationarity of historical load data.
Also, multicollinearity among exogenous variables adds to the complexities of STLF. We use data
provided by Plant-E Corp, an investor in the New York Control Area electricity market. We propose
STLF models and test them against the benchmark New York Independent System Operator (NYISO)
model; we propose three different models for STLF. The first is a hybrid model consisting of
a clustering part and a weighted Euclidean distance norm component; we name it the Cluster-WED
model. The other two models are deep learning models we shall call BiLSTM and AttnLSTM model.
Both are based on the encoder-decoder architecture. The encoder part of the BiLSTM model comprises
a bidirectional layer. The decoder is a unidirectional LSTM layer. We incorporate the attention
mechanism into the AttnLSTM model, where we assign weights to all the hidden states of the encoder
before making the next predictions in the decoder. The encoder and decoder for the AttnLSTM model
are both unidirectional LSTM layers.
Though none of our proposed models outperformed the NYISO model due to the disparities in
the input information, the results from the Cluster-WED model prove that we can perform a complex
task like STLF using simple nonlinear models. Also, the results from the BiLSTM model demonstrate
the applicability of deep learning when adapted for structured time series data. The generalization
and consistency of the BiLSTM model suggest that we could achieve competitive performance against
the benchmark once the information gap is bridged.
Keywords: Short-term load forecast, clustering, weighted Euclidean distance, deep learning, bidirectional,
unidirectional, attention.