Multivariate Prediction of Energy Time Series by Autoencoded LSTM Networks

The growing integration of renewable energy sources into modern power systems has introduced a new layer of complexity in forecasting. Unlike traditional energy sources, photovoltaic systems are inherently unstable, driven by weather conditions and seasonal variability. In this context, predicting energy production is not simply a matter of extrapolating past values, but of understanding the underlying structure of multiple interacting signals. Here, we introduce an alternative approach to this problem, combining representation learning and forecasting within a single deep learning framework. The core idea is that before predicting the future, a model should first learn how to represent the present in a meaningful way.

At the core of the methodology lies a two-stage training process built on Long Short-Term Memory (LSTM) networks. The architecture is composed of two stacked recurrent layers, each with a distinct role. The first layer is designed to act as an encoder, transforming multivariate time series into a compact representation that captures the most relevant information across variables. This phase focuses on reconstructing the input signal, forcing the network to identify the internal structure of the data. This encoding process becomes the foundation for everything that follows.

Once the encoder has learned to extract meaningful features, the model transitions to a second training phase. In this stage, the encoder is no longer updated; its parameters are frozen, preserving the learned representation. A second LSTM layer is then trained to perform the actual prediction, using the encoded information as input. The key difference is that the model no longer needs to learn both representation and prediction simultaneously. It relies on a structured internal representation that has already been optimised. This separation between encoding and prediction introduces a form of inductive bias, guiding the model toward more stable and generalizable solutions.

The predictive task is not limited to a single time series. The model operates in a multivariate setting, where several physical variables contribute to the final prediction. In the experimental setup, the system combines power output, temperature, and wind speed, each providing complementary information about the behaviour of the energy system. This integration allows the model to capture relationships that would be invisible in a univariate framework. For instance, changes in temperature or wind conditions can influence energy production in ways that are not immediately evident from the output signal alone. By encoding these interactions into a shared representation, the model effectively builds a unified view of the system dynamics.

The effectiveness of the approach is evaluated using real-world data collected from a photovoltaic plant. The experiments consider different periods of the year, including both stable and highly variable conditions, ensuring that the model is tested under realistic scenarios. The alignment between prediction and reality reflects the model’s ability to capture the underlying structure of the data rather than simply fitting surface-level trends.

Beyond performance improvements, the proposed approach suggests a broader shift in how time series forecasting models are designed. By explicitly separating representation learning from prediction, the methodology introduces a more interpretable and modular structure. The encoder becomes a tool for understanding the data, while the predictor focuses on extrapolating future behaviour. This division not only improves accuracy but also enhances the robustness of the system, making it more adaptable to new scenarios and datasets. In increasingly complex energy systems, where multiple variables interact in non-linear ways, such approaches offer a promising path forward. The ability to learn meaningful representations from raw data may prove as important as the prediction itself, reshaping the role of artificial intelligence in forecasting tasks.