Deep Learning for Time Series Forecasting: The Electric Load Case (1907.09207v1)

Abstract: Management and efficient operations in critical infrastructure such as Smart Grids take huge advantage of accurate power load forecasting which, due to its nonlinear nature, remains a challenging task. Recently, deep learning has emerged in the machine learning field achieving impressive performance in a vast range of tasks, from image classification to machine translation. Applications of deep learning models to the electric load forecasting problem are gaining interest among researchers as well as the industry, but a comprehensive and sound comparison among different architectures is not yet available in the literature. This work aims at filling the gap by reviewing and experimentally evaluating on two real-world datasets the most recent trends in electric load forecasting, by contrasting deep learning architectures on short term forecast (one day ahead prediction). Specifically, we focus on feedforward and recurrent neural networks, sequence to sequence models and temporal convolutional neural networks along with architectural variants, which are known in the signal processing community but are novel to the load forecasting one.

• The paper demonstrates that GRU-based MIMO strategies significantly improve individual load forecasting accuracy by reducing error propagation.
• It compares various architectures, revealing that simpler ERNNs can match or outperform complex gated networks for aggregated demand scenarios.
• The study highlights the need for context-specific model selection and unified metrics to advance smart grid energy management.

Deep Learning for Time Series Forecasting: The Electric Load Case

The paper "Deep Learning for Time Series Forecasting: The Electric Load Case" by Gasparin et al. provides a thorough examination of the application of deep learning architectures to the domain of electric load forecasting, a critical yet intricate element of energy management within smart grids. The exploration centers on comparing various deep learning models' performance—specifically feedforward and recurrent neural networks, sequence-to-sequence models, and temporal convolutional networks (TCNs)—applied to short-term load forecasting (STLF).

Methodology and Models

The paper implements a comprehensive assessment of the selected architectures as follows:

• Feedforward Neural Networks (FNNs) and their evolved variant, Deep Feedforward Neural Networks (DFNNs) with residual connections, serve as benchmarks for non-sequential models.
• Recurrent Neural Networks (RNNs), including Elman nets (ERNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRUs), are scrutinized for capturing temporal dependencies within electric load time series data. The analysis further distinguishes between recursive and multiple-input, multiple-output (MIMO) forecasting strategies for RNNs, probing differences in recurring vs. batch sequence predictions.
• Sequence-to-Sequence Architectures (seq2seq) are applied to leverage encoder-decoder paradigms, typically harnessed in fields such as machine translation, here examined for their adaptability in handling electric load sequences.
• Temporal Convolutional Networks (TCNs) offer an innovative perspective on sequence modeling, utilizing dilated causal convolutions to capture long-range dependencies while maintaining computational efficiency.

Experimental Evaluation

The paper employs two primary datasets to evaluate these models: individual household electric power consumption data and the GEFCom2014 dataset, which represents aggregated electric demand. This dual scenario setup aids in assessing the models' robustness across disparate levels of load predictability, from volatile household-level measurements to aggregated state-level data.

Results show that amongst RNNs, MIMO strategies, particularly those employing GRUs, furnish superior predictions for individual load scenarios, indicating a proclivity of MIMO methods to handle complex sequences due to reduced error propagation. Conversely, in the aggregated demand scenarios, simpler ERNN architectures exhibit the potential to match or surpass gated networks, emphasizing efficiency in regulated data contexts.

The comparative performance of seq2seq models trained under various conditioning paradigms suggests that alleviating exposure bias (e.g., through self-generated strategies) holds promise in enhancing predictive precision. Moreover, TCNs emerge as capable alternatives, espousing comparable efficacy to RNNs with the added advantages of improved stability and training parallelization.

Implications and Future Directions

This investigation evidences that model selection in STLF should be predicated on data characteristics, with context-specific architecture optimizations paramount. Gated recurrent networks maintain an edge in erratic load conditions, whereas simpler models suffice for regular consumption patterns. The promising outcomes of TCNs and augmentation of seq2seq methods beckon further research, particularly exploring hybrid architectures that might amalgamate TCNs' efficient temporal processing with RNNs' configurability to fortify load forecasting.

The paper asserts the necessity of unified metrics and datasets for fair performance assessments to guide future smart grid technologies. It bolsters the strategic incorporation of exogenous variables, such as weather, to amplify model accuracy—highlighting an area ripe for expansion in industrial applications. This comprehensive evaluation illuminates the nuanced landscape of deep learning for electric forecasting, establishing a groundwork for continued innovation within smart grid infrastructures.