Matteo De Felice

Short-Term Load Forecasting with Neural Network Ensembles: A Comparative Study [Application Notes]

Load Forecasting plays a critical role in the management, scheduling and dispatching operations in power systems, and it concerns the prediction of energy demand in different time spans. In future electric grids, to achieve a greater control and flexibility than in actual electric grids, a reliable forecasting of load demand could help to avoid dispatch problems given by unexpected loads, and give vital information to make decisions on energy generation and purchase, especially market-based dynamic pricing strategies. Furthermore, accurate prediction would have a significant impact on operation management, e.g. preventing overloading and allowing an efficient energy storage.

Robust assessment of the expansion and retreat of Mediterranean climate in the 21st century

The warm-temperate regions of the globe characterized by dry summers and wet winters (Mediterranean climate; MED) are especially vulnerable to climate change. The potential impact on water resources, ecosystems and human livelihood requires a detailed picture of the future changes in this unique climate zone. Here we apply a probabilistic approach to quantitatively address how and why the geographic distribution of MED will change based on the latest-available climate projections for the 21st century. Our analysis provides, for the first time, a robust assessment of significant northward and eastward future expansions of MED over both the Euro-Mediterranean and western North America. Concurrently, we show a significant 21st century replacement of the equatorward MED margins by the arid climate type. Moreover, future winters will become wetter and summers drier in both the old and newly established MED zones. Should these projections be realized, living conditions in some of the most densely populated regions in the world will be seriously jeopardized.

Combining Back-Propagation and Genetic Algorithms to Train Neural Networks for Ambient Temperature Modeling in Italy

This paper presents a hybrid approach based on soft computing techniques in order to estimate ambient temperature for those places where such datum is not available. Indeed, we combine the Back-Propagation (BP) algorithm and the Simple Genetic Algorithm (GA) in order to effectively train neural networks in such a way that the BP algorithm initialises a few individuals of the GAs population. Experiments have been performed over all the available Italian places and results have shown a remarkable improvement in accuracy compared to the single and traditional methods.

Design of Robust Space Trajectories

In this paper we apply a novel black-box optimisation algorithm to the Global Trajectory optimisation Problem provided by the European Space Agency (ESA). The proposed algorithm, called SAGES, has been applied to instances of seven trajectory design problems, comparing it with the known best solutions. The numerical results show clear improvements on the majority of the problems and, in order to investigate deeply the problems, a sensitivity and solutions robustness analysis has been performed, measuring the influence of each single variable to the objective function.

Neural networks ensembles for short-term load forecasting

This paper proposes a new approach for short-term load forecasting based on neural networks ensembling methods. A comparison between traditional statistical linear seasonal model and ANN-based models has been performed on the real-world building load data, considering the utilisation of external data such as the day of the week and building occupancy data. The selected models have been compared to the prediction of hourly demand for the electric power up to 24 hours for a testing week. Both neural networks ensembles achieved lower average and maximum errors than other models. Experiments showed how the introduction of external data had helped the forecasting.

Start-Up optimisation of a combined cycle power plant with multiobjective evolutionary algorithms

In this paper we present a study of the application of Evolutionary Computation methods to the optimisation of the start-up of a combined cycle power plant. We propose a multiobjective approach considering different objectives for the optimisation in order to reduce the pollution emissions and to maximise the efficiency of the plant. We compare a multiobjective evolutionary algorithm (NSGA-II) with 2 and 5 objectives on a software simulator and then we use different metrics to measure the performances. We show that NSGA-II algorithm is able to provide a set of solutions, defined as Pareto Front, that represent the best trade-off on the different objectives among those the decision maker can choose.

Deterministic and Stochastic Approaches for Day-Ahead Solar Power Forecasting

Photovoltaic (PV) power forecasting has the potential to mitigate some of effects of resource variability caused by high solar power penetration into the electricity grid. Two main methods are currently used for PV power generation forecast: (i) a deterministic approach that uses physics-based models requiring detailed PV plant information and (ii) a data-driven approach based on statistical or stochastic machine learning techniques needing historical power measurements. The main goal of this work is to analyze the accuracy of these different approaches. Deterministic and stochastic models for dayahead PV generation forecast were developed, and a detailed error analysis was performed. Four years of site measurements were used to train and test the models. Numerical weather prediction (NWP) data generated by the weather research and forecasting (WRF) model were used as input. Additionally, a new parameter, the clear sky performance index, is defined. This index is equivalent to the clear sky index for PV power generation forecast, and it is here used in conjunction to the stochastic and persistence models. The stochastic model not only was able to correct NWP bias errors but it also provided a better irradiance transposition on the PV plane. The deterministic and stochastic models yield day-ahead forecast skills with respect to persistence of 35% and 39%, respectively.

Evolving Complex Neural Networks

Complex networks like the scale-free model proposed by Barabasi-Albert are observed in many biological systems and the application of this topology to artificial neural network leads to interesting considerations. In this paper, we present a preliminary study on how to evolve neural networks with complex topologies. This approach is utilized in the problem of modeling a chemical process with the presence of unknown inputs (disturbance). The evolutionary algorithm we use considers an initial population of individuals with differents scale-free networks in the genotype and at the end of the algorithm we observe and analyze the topology of networks with the best performances. Experimentation on modeling a complex chemical process shows that performances of networks with complex topology are similar to the feed-forward ones but the analysis of the topology of the most performing networks leads to the conclusion that the distribution of input node information affects the network performance (modeling capability).

Dynamical Modeling and Parameter Identification of Seismic Isolation Systems by Evolution Strategies

An application of Evolution Strategies (ESs) to the dynamic identification of hybrid seismic isolation systems is presented. It is shown how ESs are highly effective for the optimisation of the test problem defined in previous work for methodology validation. The acceleration records of a number of dynamic tests performed on a seismically isolated building are used as reference data for the parameter identification. The application of CMA-ES to a previously existing model considerably improves previous results but at the same time reveals limitations of the model. To investigate the problem three new mechanical models with higher number of parameters are developed. The application of CMA-ES to the best designed model allows improvements of up to 79% compared to the solutions previously available in literature.

Soft Computing Techniques for Internet Backbone Traffic Anomaly Detection

The detection of anomalies and faults is a fundamental task for different fields, especially in real cases like LAN networks and the Internet. We present an experimental study of anomaly detection on a simulated Internet backbone network based on neural networks, particle swarms, and artificial immune systems.