This study presents a novel neuroevolutionary approach for efficient automatic discovery of optimal convolutional architectures. This is, to the best of our knowledge, the first resource-aware multi-objective neuroevolutionary architecture search (NAS) method for semantic segmentation. Towards this goal, we first address one of the main caveats in NAS, their high computational cost, with a new technique we term progressive stratified sampling (PSS). This strategy is specifically devised to speed up the search by an order of magnitude with a minimal sacrifice in performance. We demonstrate the proposed approach in a range of computer vision problems including CIFAR10 and Cityscapes datasets. Results show that the new method generates state-of-the-art competitive architectures at a significantly lower computational cost; and thus, it outperforms previous approaches in terms of balancing performance and resources. In addition, we open-source this resource-aware, multi-objective approach, providing dataset loaders for numerous problems. In short, our framework is designed to be a stepping stone towards a more general, efficient and effective multi-objective methodology for NAS.

X-ray screening is crucial for ensuring safety and security in crowded public areas. However, X-ray operators are often overwhelmed by the sheer amount of potential threats to assess; thus, current computer vision-aided systems are designed to alleviate these workloads. In this study, we focus on a key, unresolved challenge for developing such automatic X-ray screening systems: the direct application of existing avant garde computer vision approaches does not necessarily yield satisfactory results in the X-ray medium, hindering the effectiveness of current screening systems. To overcome this drawback, we propose a novel automated machine learning (AutoML) multi-objective approach for neural architecture search (NAS) for concealed weapon detection (MEOW). We benchmark MEOW with the state-of-the-art in two comprehensive scenarios in threat identification: SIXray (a popular, massive X-ray dataset) and Residuals (a proprietary, unpublished dataset provided by our industry partners). MEOW consist of the coalescence of two new components: First, we design a heuristic technique to strongly reduce the high computational cost of neuroevolutionary search while preserving a high performance such that it can be effectively used in real-time industrial settings. Second, we devise a novel ensemble approach for combining multiple discovered architectures simultaneously. Leveraging these two characteristics, MEOW outperforms the state-of-the-art while keeping the NAS overhead to a minimum. More broadly, our results suggest that AutoML has a strong potential for security applications.

We describe our approach to the Western Power Distribution (WPD) Presumed Open Data (POD) 6 MWh battery storage capacity forecasting competition, in which we finished second. The competition entails two distinct forecasting aims to maximise the daily evening peak reduction and using as much solar photovoltaic energy as possible. For the latter, we combine a Bayesian (MCMC) linear regression model with an average generation distribution. For the former, we introduce a new error metric that allows even a simple weighted average combined with a simple linear regression model to score very well using the competition performance metric.

In this paper, we present a novel neuroevolutionary method to identify the architecture and hyperparameters of convolutional autoencoders. Remarkably, we used a hypervolume indicator in the context of neural architecture search for autoencoders, for the first time to our current knowledge. Results show that images were compressed by a factor of more than 10, while still retaining enough information to achieve image classification for the majority of the tasks. Thus, this new approach can be used to speed up the AutoML pipeline for image compression.

While social networks can provide an ideal platform for upto-date information from individuals across the world, it has also proved to be a place where rumours fester and accidental or deliberate misinformation often emerges. In this article, we aim to support the task of making sense from social media data, and specifically, seek to build an autonomous message-classifier that filters relevant and trustworthy information from Twitter. For our work, we collected about 100 million public tweets, including users’ past tweets, from which we identified 72 rumours (41 true, 31 false). We considered over 80 trustworthiness measures including the authors’ profile and past behaviour, the social network connections (graphs), and the content of tweets themselves. We ran modern machine-learning classifiers over those measures to produce trustworthiness scores at various time windows from the outbreak of the rumour. Such time-windows were key as they allowed useful insight into the progression of the rumours. From our findings, we identified that our model was significantly more accurate than similar studies in the literature. We also identified critical attributes of the data that give rise to the trustworthiness scores assigned. Finally we developed a software demonstration that provides a visual user interface to allow the user to examine the analysis.

Distribution network operators (DNOs) are increasingly concerned about the impact of low carbon technologies on the low voltage (LV) networks. More advanced metering infrastructures provide numerous opportunities for more accurate load flow analysis of the LV networks. However, such data may not be readily available for DNOs and in any case is likely to be expensive. Modelling tools are required which can provide realistic, yet accurate, load profiles as input for a network modelling tool, without needing access to large amounts of monitored customer data. In this paper we outline some simple methods for accurately modelling a large number of unmonitored residential customers at the LV level. We do this by a process we call buddying, which models unmonitored customers by assigning them load profiles from a limited sample of monitored customers who have smart meters. Hence the presented method requires access to only a relatively small amount of domestic customers' data. The method is efficiently optimised using a genetic algorithm to minimise a weighted cost function between matching the substation data and the individual mean daily demands. Hence we can show the effectiveness of substation monitoring in LV network modelling. Using real LV network modelling, we show that our methods perform significantly better than a comparative Monte Carlo approach, and provide a description of the peak demand behaviour.

While social networks can provide an ideal platform for up-to-date information from individuals across the world, it has also proved to be a place where rumours fester and accidental or deliberate misinformation often emerges. In this article, we aim to support the task of making sense from social media data, and specifically, seek to build an autonomous message-classifier that filters relevant and trustworthy information from Twitter. For our work, we collected about 100 million public tweets, including users’ past tweets, from which we identified 72 rumours (41 true, 31 false). We considered over 80 trustworthiness measures including the authors’ profile and past behaviour, the social network connections (graphs), and the content of tweets themselves. We ran modern machine-learning classifiers over those measures to produce trustworthiness scores at various time windows from the outbreak of the rumour. Such time-windows were key as they allowed useful insight into the progression of the rumours. From our findings, we identified that our model was significantly more accurate than similar studies in the literature. We also identified critical attributes of the data that give rise to the trustworthiness scores assigned. Finally we developed a software demonstration that provides a visual user interface to allow the user to examine the analysis.

We study the relationship between the sentiment levels of Twitter users and the evolving network structure that the users created by @-mentioning each other. We use a large dataset of tweets to which we apply three sentiment scoring algorithms, including the open source SentiStrength program. Specifically we make three contributions. Firstly, we find that people who have potentially the largest communication reach (according to a dynamic centrality measure) use sentiment differently than the average user: for example, they use positive sentiment more often and negative sentiment less often. Secondly, we find that when we follow structurally stable Twitter communities over a period of months, their sentiment levels are also stable, and sudden changes in community sentiment from one day to the next can in most cases be traced to external events affecting the community. Thirdly, based on our findings, we create and calibrate a simple agent-based model that is capable of reproducing measures of emotive response comparable with those obtained from our empirical dataset.

Clustering methods are increasingly being applied to residential smart meter data, which provides a number of important opportunities for distribution network operators (DNOs) to manage and plan low-voltage networks. Clustering has a number of potential advantages for DNOs, including the identification of suitable candidates for demand response and the improvement of energy profile modeling. However, due to the high stochasticity and irregularity of household-level demand, detailed analytics are required to define appropriate attributes to cluster. In this paper, we present in-depth analysis of customer smart meter data to better understand the peak demand and major sources of variability in their behavior. We find four key time periods, in which the data should be analyzed, and use this to form relevant attributes for our clustering. We present a finite mixture model-based clustering, where we discover ten distinct behavior groups describing customers based on their demand and their variability. Finally, using an existing bootstrap technique, we show that the clustering is reliable. To the authors' knowledge, this is the first time in the power systems literature that the sample robustness of the clustering has been tested.

A w-local permutation is a permutation π on {1, . . . , n} that displaces each element by a distance of at most w. We define the minimum cost local permutation (MCLP) problem, where there is a cost associated with each choice of π(i) for i = 1, . . . , n and the goal is to find the w-local permutation of minimum total cost. The MCLP problem generalises the problem of computing the adjusted error, a measure of the similarity of two household-level smart meter energy profiles. Existing work has reduced MCLP to the assignment problem, which can be solved in O(n) time using the Hungarian algorithm. We prove a reduction of the MCLP problem to that of computing the shortest path in a directed acyclic graph. This yields an algorithm that, for fixed w, solves MCLP in O(n) time. Analysis of running times for adjusted error computations on a real smart meter data set confirms that the new method is far faster in practice. Further, we study N -MCLP, a generalisation of MCLP where N permutations are chosen simultaneously to minimise the associated total cost. N -MCLP generalises the problem of computing an appropriate “average” of N smart meter profiles with respect to the adjusted error measure. As in the MCLP case, we prove a reduction to the problem of computing the shortest path in a directed acyclic graph. We apply the resulting algorithm to a smart meter data set, obtaining improved forecasts for household-level energy consumption.

The Distribution Network Operators (DNOs) role is becoming more difficult as electric vehicles and electric heating penetrate the network, increasing the demand. As a result it becomes harder for the distribution networks infrastructure to remain within its operating constraints. Energy storage is a potential alternative to conventional network reinforcement such as upgrading cables and transformers. The research presented here in this paper shows that due to the volatile nature of the LV network, the control approach used for energy storage has a significant impact on performance. This paper presents and compares control methodologies for energy storage where the objective is to get the greatest possible peak demand reduction across the day from a pre-specified storage device. The results presented show the benefits and detriments of specific types of control on a storage device connected to a single phase of an LV network, using aggregated demand profiles based on real smart meter data from individual homes. The research demonstrates an important relationship between how predictable an aggregation is and the best control methodology required to achieve the objective.

We present a refined parametric model for forecasting electricity demand which performed particularly well in the recent Global Energy Forecasting Competition (GEFCom 2012). We begin by motivating and presenting a simple parametric model, treating the electricity demand as a function of the temperature and day of the data. We then set out a series of refinements of the model, explaining the rationale for each, and using the competition scores to demonstrate that each successive refinement step increases the accuracy of the model’s predictions. These refinements include combining models from multiple weather stations, removing outliers from the historical data, and special treatments of public holidays.

As low carbon technologies become more pervasive, distribution network operators are looking to support the expected changes in the demands on the low voltage networks through the smarter control of storage devices. Accurate forecasts of demand at the individual household-level, or of small aggregations of households, can improve the peak demand reduction brought about through such devices by helping to plan the most appropriate charging and discharging cycles. However, before such methods can be developed, validation measures which can assess the accuracy and usefulness of forecasts of the volatile and noisy household-level demand are required. In this paper we introduce a new forecast verification error measure that reduces the so-called “double penalty” effect, incurred by forecasts whose features are displaced in space or time, compared to traditional point-wise metrics, such as the Mean Absolute Error, and p-norms in general. The measure that we propose is based on finding a restricted permutation of the original forecast that minimises the point-wise error, according to a given metric. We illustrate the advantages of our error measure using half-hourly domestic household electrical energy usage data recorded by smart meters, and discuss the effect of the permutation restriction.

We present an efficient graph-based algorithm for quantifying the similarity of household-level energy use profiles, using a notion of similarity that allows for small time-shifts when comparing profiles. Experimental results on a real smart meter data set demonstrate that in cases of practical interest our technique is far faster than the existing method for computing the same similarity measure. Having a fast algorithm for measuring profile similarity improves the efficiency of tasks such as clustering of customers and cross-validation of forecasting methods using historical data. Furthermore, we apply a generalisation of our algorithm to produce substantially better household-level energy use forecasts from historical smart meter data.

We analyse spectral properties of 1671 galaxies from the Stromlo-APM Survey, selected to have 15 less than or equal to b(J) less than or equal to 17.15 and having a mean redshift z = 0.05. This is a representative local sample of field galaxies, so the global properties of the galaxy population provide a comparative point for analysis of more distant surveys. We measure H alpha, [O II] lambda 3727, [S II] lambda lambda 6716, 6731, [N II] lambda 6583 and [O I] lambda 6300 equivalent widths and the D-4000 break index. The 5-Angstrom-resolution spectra use an 8-arcsec slit, which typically covers 40-50 per cent of the galaxy area. We find no evidence for systematic trends depending on the fraction of galaxy covered by the slit, and further analysis suggests that our spectra are representative of integrated galaxy spectra.We classify spectra according to their H alpha emission, which is closely related to massive star formation. Overall we find that 61 per cent of galaxies are H alpha emitters with rest-frame equivalent widths EW(H alpha) greater than or similar to 2 Angstrom. The emission-line galaxy (ELG) fraction is smaller than seen in the Canada-France Redshift Survey (CFRS) at z = 0.2, and is consistent with a rapid evolution of H alpha luminosity density. The ELG fraction and EW(H alpha) increase at fainter absolute magnitudes, smaller projected area and smaller D-4000. In the local Universe, faint, small galaxies are dominated by star formation activity, while bright, large galaxies are more quiescent. This picture of the local Universe is quite different from that of the distant one; bright galaxies appear to show rapidly increasing activity as one moves further back in time.We find that the ratio [N II] lambda 6583/H alpha is anticorrelated with EW(H alpha), and that the value of 0.5 commonly used to remove the [N II] contribution from blended H alpha + [N II] lambda lambda 6548, 6583 applies only for samples with an EW distribution similar to that seen at low redshift. We show that the [O II], [N II], [S II] and H alpha EWs are correlated, but with large dispersions (similar to 50 per cent) owing to the diversity of galaxy contents sampled. Our [O II]-H alpha relation is similar to the one derived by Kennicutt, but is 10 per cent higher at 1 sigma significance. We show that this relation is not valid for distant, strong [O II] emitters with blue colours, which are more numerous than locally. This relation would overestimate the individual star formation rate by similar to 50 per cent for these kinds of galaxies. We find that, on average, luminous blue ELGs are likely to be enhanced in nitrogen abundance. This suggests that in faint, low-mass, late-type ELGs nitrogen is a primary element, whereas in brighter, more massive galaxies nitrogen comes from a secondary source. We also find that 4 per cent of early-type galaxies show star formation activity; this fraction seems to increase at higher redshifts.

Recent work introduced a novel time-permuting error measure for forecasts of half-hourly, household-level energy demand, designed to reward forecasts which predict extremes (spikes) in demand at approximately the right times, albeit perhaps slightly early or late. In many applications such as smart storage control, such forecasts are preferable to those that predict no spikes at all. Building on that idea, we make three contributions. First we introduce a probabilistic framework to estimate error distributions for actuals about forecasts, using the time-permuting error measure. The framework includes a variable discount for older, possibly less relevant data. Second we employ this framework to derive conditions to be satisfied by the optimal forecast under the time-permuting error measure. In turn this requires a mixture of discrete (non derivative) optimisation and calculus to condition forecasts on available historical observations. Finally we demonstrate the usefulness of our framework by using it to forecast the daily energy demand profiles for a large number of domestic energy customers. In particular we illustrate how such customers might be classified according to the relative forecastability of their behaviour and the corresponding need for different amounts of history to achieve such forecasts.