(623f) Autoencoder Based Dimensionality Reduction to Select Representative Periods for Energy System Planning Models

Deeply decarbonized power and energy systems are expected to witness drastically different supply and demand dynamics as compared to the current energy system due to several factors including: a) increasing variability in primary energy supply, mainly from variable renewable resources, b) growing role for demand-side participation such as flexible charging of electric vehicles, and c) growing role for energy storage devices that couple energy system operations over multiple time-scales ranging from hours to weeks. These factors have led to growing recognition of the importance of incorporating increasing operational detail in capacity expansion models (CEM) used for long-term energy infrastructure planning, while maintaining their computational tractability. One commonly used strategy for formulating tractable CEMs involves modeling system operations over representative periods, such as days or weeks, so as to approximate outcomes of the full-space CEM. In these cases, representative period selection (RPS) methods are often based on direct application of clustering algorithms, such as k-means. However, such methods are limited by the declining performance of clustering algorithms like k-means with increasing dimensionality of input data. Moreover, this factor also limits the ability to consider new features in the clustering process that can distinguish the importance of various input data on CEM outcomes.

Here, we propose a new class of RPS methods for energy system planning models that address the limitations of conventional methods by incorporating: a) dimensionality reduction using auto-encoders prior to clustering and b) using estimated outputs as additional features in the RPS. The autoencoder consists of a 3-layer neural network (encoder) to reduce the dimension of the original data prior to the clustering process and a 2-layer neural network (decoder) to retrieve the representative periods identified from the clustering process in the original dimension. The autoencoder improves the performance of the clustering algorithm, but also facilitates using additional features such as estimated outputs produced from CEM evaluation over disjoint periods in the input data set in parallel (e.g. 1 week or 1 day). We propose three alternative RPS methods using dimensionality reduction, with two methods considering input and estimated output features in the clustering process using one and three auto-encoders, respectively. These three methods are compared against conventional RPS methods without dimensionality reduction in terms of the ability of the corresponding reduced-space CEM models to reproduce outcomes of the full-space CEM models. Extensive numerical experimentation across 1-bus, 3-bus and 8-bus electric power networks defined using data from the Texas region demonstrate that the proposed RPS method do not add a significant computational burden and can better reproduce full-space CEM outcomes - capacity, generation, cost and non-served energy - compared to conventional RPS methods. Moreover, one of the RPS methods leads to smaller magnitude of error in reproducing full-space CEM outcomes while using half the number of representative periods as conventional RPS methods (4 vs. 8 weeks), which points to the potential for speeding up CEM evaluation enabled by the method.