A Blockwise Embedding for Multi-Day-Ahead Prediction of Energy Time Series by Randomized Deep Neural Networks

Forecasting energy time series has always been a delicate balance between accuracy and computational cost. As systems become more complex and data grows in volume, traditional deep learning models, while powerful, often require significant training effort and careful tuning. The work described here moves in a different direction, exploring how part of this complexity can be reduced without sacrificing predictive performance. At the core of the proposed solution lies a rethinking of both how temporal data is structured and how neural networks are trained. Instead of relying entirely on fully trained deep architectures, the model introduces controlled randomization within a hybrid structure, combining convolutional processing and recurrent memory in a more efficient framework.

A key innovation concerns the way time series are embedded before entering the learning model. Rather than treating past observations as a simple sequence of lagged values, the data is reorganized into blocks corresponding to entire days. Each input is constructed as a matrix where rows represent different past days and columns represent samples within those days. This blockwise embedding transforms a one-dimensional time series into a two-dimensional structure that can be processed by convolutional layers. The result is a representation that captures not only temporal evolution but also recurring daily patterns, making it particularly suitable for energy data characterized by strong periodicity. This shift is not merely a format change. It enables the model to detect correlations both within a single day and across multiple days, effectively enriching the information available for prediction.

Once the data is embedded into this structured format, it is processed by a deep architecture that combines a two-dimensional convolutional layer with an LSTM network. The convolutional layer extracts features from the blockwise representation, while the LSTM captures temporal dependencies and generates the final predictions. The distinctive aspect of the approach lies in the randomization of the convolutional layer. Instead of learning all parameters through backpropagation, part of the network is initialized randomly and kept fixed during training. This significantly reduces the number of trainable parameters, leading to faster training times while maintaining competitive accuracy. The diagram on page 3 clearly shows this pipeline, where the convolutional feature maps are flattened and passed to the LSTM layer, which then produces a full-day prediction in a single step.

Another important contribution is the shift from point-wise forecasting to block prediction. Instead of predicting one time step at a time, the model outputs an entire day of values simultaneously. This is achieved by formulating the problem as a regression task over a vector of future samples, rather than a single scalar value. This approach aligns well with real-world energy management scenarios, where decisions are often made on daily or multi-day horizons. It also avoids the accumulation of errors that typically arise in recursive forecasting strategies.

The proposed method is evaluated using real photovoltaic power data collected over an entire year, with hourly resolution. The experimental setup considers both short-term and multi-day forecasting scenarios, reflecting realistic operational needs in energy systems. The results show that the proposed architecture achieves lower prediction errors than a standard LSTM model in most cases, particularly in multi-day forecasting scenarios. At the same time, the randomized version of the model demonstrates a clear advantage in terms of training efficiency. An additional insight emerges from the training time analysis: the model can be retrained in a matter of seconds on standard hardware, making it suitable for real-time or near-real-time applications.

The broader significance of this work lies in its perspective on deep learning design. Rather than pushing toward increasingly complex and fully trainable architectures, it explores a middle ground where randomization and structure are used to control complexity. The combination of blockwise embedding, convolutional feature extraction, and recurrent modelling suggests a general strategy for handling structured temporal data. At the same time, the introduction of partially randomized layers highlights a promising direction for reducing computational cost without compromising model effectiveness. In this view, the future of time series forecasting may not depend solely on deeper networks, but on smarter ways of representing data and distributing learning across the architecture.