AutoPV: Automatically Design Your Photovoltaic Power Forecasting Model

TL;DR

The paper addresses the challenge of selecting and tuning predictive architectures for photovoltaic power forecasting (PVPF) by introducing AutoPV, a neural architecture search (NAS) framework tailored to time series forecasting in the PVPF domain. AutoPV builds a four-stage search space with 12 parameters to explore about 2.5e6 architectures and uses the MoBananas multi-objective Bayesian algorithm to balance forecast accuracy and model size. Evaluated on a Daqing PV Station dataset, AutoPV outperforms a broad set of baselines across Task 1 (historical data only) and Task 2 (with future weather), with notable gains when future weather is leveraged (up to ~9.88% MAE improvement against strong baselines). The work demonstrates the feasibility and value of applying NAS to TSF problems, offering a practical tool for non-experts and industry to automatically design effective PVPF models and paving the way for broader NAS applications in time series forecasting.

Abstract

Photovoltaic power forecasting (PVPF) is a critical area in time series forecasting (TSF), enabling the efficient utilization of solar energy. With advancements in machine learning and deep learning, various models have been applied to PVPF tasks. However, constructing an optimal predictive architecture for specific PVPF tasks remains challenging, as it requires cross-domain knowledge and significant labor costs. To address this challenge, we introduce AutoPV, a novel framework for the automated search and construction of PVPF models based on neural architecture search (NAS) technology. We develop a brand new NAS search space that incorporates various data processing techniques from state-of-the-art (SOTA) TSF models and typical PVPF deep learning models. The effectiveness of AutoPV is evaluated on diverse PVPF tasks using a dataset from the Daqing Photovoltaic Station in China. Experimental results demonstrate that AutoPV can complete the predictive architecture construction process in a relatively short time, and the newly constructed architecture is superior to SOTA predefined models. This work bridges the gap in applying NAS to TSF problems, assisting non-experts and industries in automatically designing effective PVPF models.

Figures (8)

• Figure 1: The pipeline for leveraging AutoPV in PVPF tasks involves two distinct scenarios: PVPF Task 1, which utilizes only historical data series as input, and PVPF Task 2, which incorporates additional unstable future weather information. Zero padding is used to fill in the missing dimension of the future data. The AutoPV system comprises three main components: a search space, an evaluator, and a search strategy. Given a specific task type and dataset, AutoPV automatically searches for an optimal model architecture. This architecture can then directly perform the forecasting task, as the training process is completed during the architecture search.
• Figure 2: A minor modification in the core predictive structure of AutoPV. In typical TSF models, the forecasting results for all features are generated first, then the PV power sequence is selected. In contrast, the AutoPV framework adds a fully connected layer to aggregate all forecasting results of different features and generate the final PV power forecast.
• Figure 3: The detailed process of neural architecture search and construction. The FFM and FFT parameters determine the final feature set. The DGM, SM, and FAM parameters conduct data augmentation on the original data. Time embedding features can be optionally added at this stage. The FEM parameter extracts different types of information from the processed data. As an example, $\mathbf{FEM}_3$ decomposes the data into trend and seasonal parts. The seasonal part is then input into the core predictive structure, which is decided by the CPS, LN, and HS parameters. The trend part is embedded by a fully connected (FC) layer and added to the output of the core predictive structure. For other $\mathbf{FEM}_x$ methods that output only a single tensor, we process the tensor as the pipeline of the seasonal part. Finally, the loss value is calculated using the prediction output and the ground truth, which serves as the evaluation score. The BS, OP, and LR parameters decide the training settings of the overall architecture.
• Figure 4: Search performance of different PVPF tasks.
• Figure 5: The average search time of each iteration and the search wall time for finding the optimal architecture across different PVPF tasks.