Two-stage dynamic management in energy communities using a decision system based on elastic net regularization

Time series analysis represents one of the most relevant challenges in artificial intelligence applied to real-world systems. Sequential data appear everywhere: in energy systems, finance, industrial sensors, communication networks, and environmental monitoring applications. However, when these sequences become multivariate and involve nonlinear relationships among many variables, traditional methods often struggle to capture the underlying dynamics. In this work, we address this difficulty by introducing a strategy that combines structured data representations with advanced deep learning models. The goal is to improve the ability of predictive models to interpret the complexity of multivariate time series by simultaneously exploiting spatial and temporal information embedded in the data.

One of the most innovative aspects concerns the way time series are represented before being processed by a neural network. Instead of treating each variable as a simple independent sequence over time, the data are reorganised into a two-dimensional structure that relates variables and time steps. This transformation enables the reinterpretation of time series as information maps. In this form, the data can be analysed using two-dimensional convolution operations, techniques typically employed in image recognition but particularly effective in identifying complex patterns among correlated variables. The idea behind this choice is that many relationships present in time series do not emerge only along the temporal dimension, but also from interactions among different variables observed within the same time interval. Representing these interactions in a two-dimensional format allows the model to identify correlations that would otherwise remain hidden.

Once this structured representation of the data is built, the learning architecture leverages two fundamental components of modern deep learning. The first layer consists of a bidimensional convolutional neural network designed to automatically extract relevant features from the maps generated from the time series. Through convolutional filters, the model identifies local patterns and correlations among variables that help describe the behaviour of the system. The information extracted during this phase is then processed by a recurrent neural network of the LSTM type, a model specifically designed to analyse temporal sequences and capture long-term dependencies. Thanks to its internal memory mechanisms, the LSTM is able to retain relevant information even when relationships among events occur across extended time horizons. The resulting architecture therefore combines two complementary capabilities: on the one hand, the ability to recognise complex patterns among different variables, and on the other hand, the ability to model the temporal evolution of the observed phenomenon.

To verify the effectiveness of the method, the system is applied to real data collected from a photovoltaic plant together with associated meteorological measurements. The dataset includes numerous environmental and operational variables, such as temperature, wind speed, wind direction, humidity, and atmospheric pressure, alongside the power generated by the plant. The analysis focuses on forecasting energy production in the following hours, a problem that is particularly relevant for power grid management and for integrating renewable energy sources into the energy system. The performance of the model is compared with architectures based solely on traditional LSTM networks. The results show that the combination of bidimensional data representation, convolution, and recurrent modelling allows the system to capture more effectively the relationships between meteorological variables and energy production, reducing prediction errors compared to conventional approaches.

Beyond the specific results in the energy domain, the most significant contribution lies in the way sequential data are interpreted. Transforming time series into bidimensional structures opens the possibility of applying tools originally developed for computer vision to problems traditionally associated with temporal analysis. This type of integration between different techniques represents a promising direction for the evolution of artificial intelligence applied to complex systems. When data coming from sensors, environmental phenomena, and digital infrastructures are analysed with models capable of capturing their deep interactions, it becomes possible to transform large volumes of measurements into operational knowledge. In this perspective, artificial intelligence no longer limits itself to predicting the behaviour of a single variable, but becomes a tool for understanding the dynamic structure of real-world systems and supporting informed decision-making in highly complex environments.