Abstract Wind and solar energy forecasting have become crucial for the inclusion of renewable energy in electrical power systems. Although most works have focused on point prediction, it is currently becoming important to also estimate the forecast uncertainty. With regard to forecasting methods, deep neural networks have shown good performance in many fields. However, the use of these networks for comparative studies of probabilistic forecasts of renewable energies, especially for regional forecasts, has not yet received much attention. The aim of this article is to study the performance of deep networks for estimating multiple conditional quantiles on regional renewable electricity production and compare them with widely used quantile regression methods such as the linear, support vector quantile regression, gradient boosting quantile regression, natural gradient boosting and quantile regression forest methods. A grid of numerical weather prediction variables covers the region of interest. These variables act as the predictors of the regional model. In addition to quantiles, prediction intervals are also constructed, and the models are evaluated using different metrics. These prediction intervals are further improved through an adapted conformalized quantile regression methodology. Overall, the results show that deep networks are the best performing method for both solar and wind energy regions, producing narrow prediction intervals with good coverage.

Abstract Machine learning is routinely used to forecast solar radiation from inputs, which are forecasts of meteorological variables provided by numerical weather prediction (NWP) models, on a spatially distributed grid. However, the number of features resulting from these grids is usually large, especially if several vertical levels are included. Principal Components Analysis (PCA) is one of the simplest and most widely-used methods to extract features and reduce dimensionality in renewable energy forecasting, although this method has some limitations. First, it performs a global linear analysis, and second it is an unsupervised method. Locality Preserving Projection (LPP) overcomes the locality problem, and recently the Linear Optimal Low-Rank (LOL) method has extended Linear Discriminant Analysis (LDA) to be applicable when the number of features is larger than the number of samples. Supervised Nonnegative Matrix Factorization (SNMF) also achieves this goal extending the Nonnegative Matrix Factorization (NMF) framework to integrate the logistic regression loss function. In this article we try to overcome all these issues together by proposing a Supervised Local Maximum Variance Preserving (SLMVP) method, a supervised non-linear method for feature extraction and dimensionality reduction. PCA, LPP, LOL, SNMF and SLMVP have been compared on Global Horizontal Irradiance (GHI) and Direct Normal Irradiance (DNI) radiation data at two different Iberian locations: Seville and Lisbon. Results show that for both kinds of radiation (GHI and DNI) and the two locations, SLMVP produces smaller MAE errors than PCA, LPP, LOL, and SNMF, around 4.92% better for Seville and 3.12% for Lisbon. It has also been shown that, although SLMVP, PCA, and LPP benefit from using a non-linear regression method (Gradient Boosting in this work), this benefit is larger for PCA and LPP because SMLVP is able to perform non-linear transformations of inputs.

Deep neural networks (DNN) are becoming increasingly relevant for probabilistic forecasting because of their ability to estimate prediction intervals (PIs). Two different ways for estimating PIs with neural networks stand out: quantile estimation for posterior PI construction and direct PI estimation. The former first estimates quantiles , which are then used to construct PIs, while the latter directly obtains the lower and upper PI bounds by optimizing some loss functions, with the advantage that PI width is directly considered in the optimization process and thus may result in narrower intervals. In this work, two different DNN-based models are studied for direct PI estimation, and compared with DNN for quantile estimation in the context of solar and wind regional energy forecasting. The first approach is based on the recent quality-driven loss and is formulated to estimate multiple PIs with a single model. The second is a novel approach that employs hypernetworks (HN), where direct PI estimation is formulated as a multi-objective problem, returning a Pareto front of solutions that contains all possible coverage-width optimal trade-offs. This formulation allows HN to obtain optimal PIs for all possible coverages without increasing the number of network outputs or adjusting additional hyperparameters, as opposed to the first direct model. Results show that prediction intervals from direct estimation are narrower (up to 20%) than those of quantile estimation, for target coverages 70%–80% for all regions, and also 85%, 90%, and 95% depending on the region, while HN always achieves the required coverage for the higher target coverages.

Summary This article addresses two issues in solar energy forecasting from the numerical weather prediction (NWP) models using machine learning. First, we are interested in determining the relevant information for the forecasting task. With this purpose, a study has been carried out to evaluate the influence on accuracy of the number of NWP grid nodes used as input for the forecasting model, as well as their relative importance. Several machine learning (support vector machines and gradient boosting) and feature selection algorithms (linear, ReliefF, and local information analysis) have been used in this study. The second aim is to be able to predict solar energy for locations where no previous production data are available. To address this goal, an approach consisting on modeling regions in the grid is proposed. Models (aggregate models) use as input attributes the meteorological variables relevant for the region and two new inputs to identify the location of each station: the latitude and the longitude. Those models can be used to predict energy production for existing stations and for new locations, represented by latitude and longitude. Copyright © 2015 John Wiley & Sons, Ltd.

Receiver Operating Characteristics (ROC) curves represent the performance of a classifier for all possible operating conditions, i.e., for all preferences regarding the tradeoff between false positives and false negatives. The generation of a ROC curve generally involves the training of a single classifier for a given set of operating conditions, with the subsequent use of threshold-moving to obtain a complete ROC curve. Recent work has shown that the generation of ROC curves may also be formulated as a multi-objective optimization problem in ROC space: the goals to be minimized are the false positive and false negative rates. This technique also produces a single ROC curve, but the curve may derive from operating points for a number of different classifiers. This paper aims to provide an empirical comparison of the performance of both of the above approaches, for the specific case of prototype-based classifiers. Results on synthetic and real domains shows a performance advantage for the multi-objective approach.

Machine Learning techniques are routinely applied to Brain Computer Interfaces in order to learn a classifier for a particular user. However, research has shown that classification techniques perform better if the EEG signal is previously preprocessed to provide high quality attributes to the classifier. Spatial and frequency-selection filters can be applied for this purpose. In this paper, we propose to automatically optimize these filters by means of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The technique has been tested on data from the BCI-III competition, because both raw and manually filtered datasets were supplied, allowing to compare them. Results show that the CMA-ES is able to obtain higher accuracies than the datasets preprocessed by manually tuned filters.

Stacking is a widely used technique for combining classifiers and improving prediction accuracy. Early research in Stacking showed that selecting the right classifiers, their parameters and the meta-classifiers was a critical issue. Most of the research on this topic hand picks the right combinatio n of classifiers and their parameters. Instead of starting from these initial strong assumptions, our approach uses genetic algorithms to search for good Stacking configurations. Since this can lead to overfitting, one of the goals of this paper is to empirically evaluate the overall efficiency of the approach. A second goal is to compare our approach with the current best Stacking building techniques. The results show that our approach finds Stacking configurations that, in the worst case, perform as well as the best techniques, with the advantage of not having to manually set up the structure of the Stacking system.

Many classification algorithms use the concept of distance or similarity between patterns. Previous work has shown that it is advantageous to optimize general Euclidean distances (GED). In this paper, we optimize data transformations, which is equivalent to searching for GEDs, but can be applied to any learning algorithm, even if it does not use distances explicitly. Two optimization techniques have been used: a simple local search (LS) and the covariance matrix adaptation evolution strategy (CMA-ES). CMA-ES is an advanced evolutionary method for optimization in difficult continuous domains. Both diagonal and complete matrices have been considered. The method has also been extended to a quadratic non-linear transformation. Results show that in general, the transformation methods described here either outperform or match the classifier working on the original data.

Determining if a group of proteins are functionally associated among themselves is an open problem in molecular biology. Within our long term goal of applying Genetic Programming (GP) to this domain, this paper evaluates the feasibility of GP to predict if a given pair of proteins interacts. GP has been chosen because of its potential flexibility in many aspects, such as the definition of operations. In this paper, the if-unknown operation is defined, which semantically is the most appropriate in this domain for handling missing values. We have also used the Tarpeian bloat control method to decrease the computational time and the solution size. Our results show that GP is feasible for this domain and that the Tarpeian method can obtain large improvements in search efficiency and interpretability of solutions.

The Frequency Assignment Problem (FAP) is one of the key issues in the design of GSM networks (Global System for Mobile communications), and will remain important in the foreseeable future. There are many versions of FAP, most of them benchmarking-like problems. We use a formulation of FAP, developed in published work, that focuses on aspects which are relevant for real-world GSM networks. In this paper, we have designed, adapted, and evaluated several types of metaheuristic for different time ranges. After a detailed statistical study, results indicate that these metaheuristics are very appropriate for this FAP. New interference results have been obtained, that significantly improve those published in previous research.

Grammatical evolution is an evolutionary algorithm able to develop, starting from a grammar, programs in any language. Starting from the point that individual learning can improve evolution, in this paper it is proposed an extension of Grammatical evolution that looks at learning by reinforcement as a learning method for individuals. This way, it is possible to incorporate the Baldwinian mechanism to the evolutionary process. The effect is widened with the introduction of the Lamarck hypothesis. The system is tested in two different domains: a symbolic regression problem and an even parity Boolean function. Results show that for these domains, a system which includes learning obtains better results than a grammatical evolution basic system.

The Robosoccer simulator is a challenging environment, where a human introduces a team of agents into a football virtual environment. Typically, agents are programmed by hand, but it would be a great advantage to transfer human experience into football agents. The first aim of this paper is to use machine learning techniques to obtain models of humans playing Robosoccer. These models can be used later to control a Robosoccer agent. However, models did not play as smoothly and optimally as the human. To solve this problem, the second goal of this paper is to incrementally correct models by means of evolutionary techniques, and to adapt them against more difficult opponents than the ones beatable by the human.

Stacking is one of the most used techniques for combining classifiers and improves prediction accuracy. Early research in stacking showed that selecting the right classifiers, their parameters and the metaclassifiers was the main bottleneck for its use. Most of the research on this topic selects by hand the right combination of classifiers and their parameters. Instead of starting from these initial strong assumptions, our approach uses genetic algorithms to search for good stacking configurations. Since this can lead to overfitting, one of the goals of This work is to evaluate empirically the overall efficiency of the approach. A second goal is to compare our approach with current best stacking building techniques. The results show that our approach finds stacking configurations that, in the worst case, perform as well as the best techniques, with the advantage of not having to set up manually the structure of the stacking system.

A very important issue in multi-agent systems is that of adaptability to other agents, be it to cooperate or to compete. In competitive domains, the knowledge about the opponent can give any player a clear advantage. In previous work, we acquired models of another agent (the opponent) based only on the observation of its inputs and outputs (its behavior) by formulating the problem as a classification task. In this paper we extend this previous work to the RoboCup domain. However, we have found that models based on a single classifier have bad accuracy, To solve this problem, In this paper we propose to decompose the learning task into two tasks: learning the action name (i.e. kick or dash) and learning the parameter of that action. By using this hierarchical learning approach accuracy results improve, and at worst, the agent can know what action the opponent will carry out, even if there is no high accuracy on the action parameter.

In standard GP there are no constraints on the structure to evolve: any combination of functions and terminals is valid. However, sometimes GP is used to evolve structures that must respect some constraints. Instead of "ad-hoc" mechanisms, grammars can be used to guarantee that individuals comply with the language restrictions. In addition, grammars permit great flexibility to define the search space. EVOCK (Evolution of Control Knowledge) is a GP based system that learns control rules for PRODIGY, an AI planning system. EVOCK uses a grammar to constrain individuals to PRODIGY 4.0 control rule syntax. The authors describe the grammar specific details of EVOCK. Also, the grammar approach flexibility has been used to extend the control rule language utilized by EVOCK in earlier work. Using this flexibility, tests were performed to determine whether using combinations of several types of control rules for planning was better than using only the standard select type. Experiments have been carried out in the blocksworld domain that show that using the combination of types of control rules does not get better individuals, but it produces good individuals more frequently.

In this paper we describe SHAMASH, a tool for modeling and automatically optimizing Business Processes. The main features that differentiate it from most current related tools are its ability to define and use organisation standards, and functional structure, and make automatic model simulations and optimisation of them. SHAMASH is a knowledge based system, and we include a discussion on how knowledge acquisition takes place. Furthermore, we introduce a high level description of the architecture, the conceptual model, and other important modules of the system.

Nowadays, there is no doubt that machine learning techniques can be successfully applied to data mining tasks. Currently, the combination of several classifiers is one of the most active fields within inductive machine learning. Examples of such techniques are boosting, bagging and stacking. From these three techniques, stacking is perhaps the less used one. One of the main reasons for this relates to the difficulty to define and parameterize its components: selecting which combination of base classifiers to use, and which classifier to use as the meta-classifier. One could use for that purpose simple search methods (e.g. hill climbing), or more complex ones (e.g. genetic algorithms). But before search is attempted, it is important to know the properties of the search space itself. In this paper we study exhaustively the space of stacking systems that can be built by using four base learning systems: C4.5, IB1, Naive Bayes, and PART. The results that have been obtained in this paper will be useful for designing new Stacking-based algorithms and tools.

Solving problems in dynamic and heterogeneous environments where information sources change their format representation and stored data is very complex. In previous work we presented a system called MAPWeb (Multiagent Planning on the Web) that tried to solve these problems by integrating artificial intelligence planning techniques within the multiagent framework. Basically, MAPWeb allows cooperative work between planning agents and Web agents. The purpose of MAPWeb is to find solutions to travel problems. In order to give detailed solutions, MAPWeb uses information gathering techniques to retrieve travel information that is made available by many different companies. However, Web access to the information sources is quite time expensive. In this paper, we try to minimize the number of Web queries by using caching techniques based on relational databases. Experimental results show that the reduction in Web access time is quite important, while maintaining the number of solutions found.

Declarative problem solving, such as planning, poses interesting challenges for Genetic Programming (GP). There have been recent attempts to apply GP to planning that fit two approaches: (a) using GP to search in plan space or (b) to evolve a planner. In this article, we propose to evolve only the heuristics to make a particular planner more efficient. This approach is more feasible than (b) because it does not have to build a planner from scratch but can take advantage of already existing planning systems. It is also more efficient than (a) because once the heuristics have been evolved, they can be used to solve a whole class of different planning problems in a planning domain, instead of running GP for every new planning problem. Empirical results show that our approach (EvoCK) is able to evolve heuristics in two planning domains (the blocks world and the logistics domain) that improve PRODIGY4.0 performance. Additionally, we experiment with a new genetic operator --Instance-Based Crossover--that is able to use traces of the base planner as raw genetic material to be injected into the evolving population.

Multilayer feedforward neural networks with backpropagation algorithm have been used successfully in many applications. However, the level of generalization is heavily dependent on the quality of the training data. That is, some of the training patterns can be redundant or irrelevant. It has been shown that with careful dynamic selection of training patterns, better generalization performance may be obtained. Nevertheless, generalization is carried out independently of the novel patterns to be approximated. In this paper, we present a learning method that automatically selects the training patterns more appropriate to the new sample to be predicted. This training method follows a lazy learning strategy, in the sense that it builds approximations centered around the novel sample. The proposed method has been applied to three different domains: two artificial approximation problems and a real time series prediction problem. Results have been compared to standard backpropagation using the complete training data set and the new method shows better generalization abilities.