Harri Niska

Extensive evaluation of neural network models for the prediction of NO2 and PM10 concentrations, compared with a deterministic modelling system and measurements in central Helsinki

Five neural network (NN) models, a linear statistical model and a deterministic modelling system (DET) were evaluated for the prediction of urban NO2 and PM10 concentrations. The model evaluation work considered the sequential hourly concentration time series of NO2 and PM10, which were measured at two stations in central Helsinki, from 1996 to 1999. The models utilised selected traffic flow and pre-processed meteorological variables as input data. An imputed concentration dataset was also created, in which the missing values were replaced, in order to obtain a harmonised database that is well suited for the inter-comparison of models. Three statistical criteria were adopted: the index of agreement (IA), the squared correlation coefficient (R2) and the fractional bias. The results obtained with various non-linear NN models show a good agreement with the measured concentration data for NO2; for instance, the annual mean of the IA values and their standard deviations range from 0.86±0.02 to 0.91±0.01. In the case of NO2, the non-linear NN models produce a range of model performance values that are slightly better than those by the DET. NN models generally perform better than the statistical linear model, for predicting both NO2 and PM10 concentrations. In the case of PM10, the model performance statistics of the NN models were not as good as those for NO2 over the entire range of models considered. However, the currently available NN models are neither applicable for predicting spatial concentration distributions in urban areas, nor for evaluating air pollution abatement scenarios for future years.

Evolving the neural network model for forecasting air pollution time series

Abstract The modelling of real-world processes such as air quality is generally a difficult task due to both their chaotic and non-linear phenomenon and high dimensional sample space. Despite neural networks (NN) have been used successfully in this domain, the selection of network architecture is still problematic and time consuming task when developing a model for practical situation. This paper presents a study where a parallel genetic algorithm (GA) is used for selecting the inputs and designing the high-level architecture of a multi-layer perceptron model for forecasting hourly concentrations of nitrogen dioxide at a busy urban traffic station in Helsinki. In addition, the tuning of GAs parameters for the problem is considered in experimental way. The results showed that the GA is a capable tool for tackling the practical problems of neural network design. However, it was observed that the evaluation of NN models is a computationally expensive process, which set limits for the search techniques.

Use of sewer on-line total solids data in wastewater treatment plant modelling.

We describe a neural network model of a municipal wastewater treatment plant (WWTP) in which on-line total solids (TS) sewer data generated by a novel microwave sensor is used as a model input variable. The predictive performance of the model is compared with and without sewer data and with modelling with a traditional linear multiple linear regression (MLR) model. In addition, the benefits of using neural networks are discussed. According to our results, the neural network based MLP (multilayer perceptron) model provides a better estimate than the corresponding MLR model of WWTP effluent TS load. The inclusion of sewer TS data as an input variable improved the performance of the models. The results suggest that increased on-line sensing of WWTPs should be stressed and that neural networks are useful as a modelling tool due to their capability of handling the nonlinear and dynamic data of sewer and WWTP systems.

An Experimental Evaluation of ZCS-DM for the Prediction of Urban Air Quality

Understanding and forecasting urban Air Quality (AQ) is not only a multi-faceted and computationally challenging problem for machine learning algorithms, but also a difficult task for human-decision makers: the strict regulatory framework, in combination with the public demand for better information services poses the need for robust, efficient and, more importantly, understandable forecasting models. Unlike neural network or regression-based techniques traditionally used in the domain of AQ, our current approach adopts ZCS-DM – Zeroth-level Classifier System for Data Mining – for the production of a set of AQ prediction rules in the urban domain. On this basis, the aim of our experimental investigation is two-fold and includes (i) the handling of incomplete data matrices and (ii) the evaluation of ZCS-DM effectiveness against methods widely used in the target domain, namely multi-layer perceptron (MLP), support vector machines (SVM), linear discriminant analysis (LDA) and classification trees (C4.5). Overall, the obtained results reveal the potential of ZCS-DM as a data mining tool to be used for AQ forecasting, and point to its insensitivity for missing values (cf. MLP, SVM, LDA) and the understandability of produced models as its greater advantages.

Hybrid Models for Forecasting Air Pollution Episodes

Urban air pollutants have emerged as a severe problem which causes health effects and even premature deaths among sensitive groups. Therefore a warning system for air pollution episodes is widely needed to minimize negative health effects. However the forecasting of air pollution episodes has been observed to be problematic partly due their rareness and short-term nature. The research presented here aims to evaluate different neural network based models for forecasting urban air pollution (N02) hourly time series and particularly the episode peaks. The performances of three multi-layer perceptron (MLP) models namely basic MLP and two hybrid models were compared by calculating several statistical indicators. In the hybrid models evaluated here, training data set was clustered to several air quality episodes using the k-means (KM) and fuzzy c-means (FCM) algorithms and then several MLP models were applied to the clustered data, each one representing one cluster. The results showed that the hybrid models have some advantages over a basic MLP model in the forecasting air quality episodes, but the performance achieved also show that architectural issues cannot solely solve the model performance problems.

Evolving smart meter data driven model for short-term forecasting of electric loads

Short-term forecasting of electric loads is an essential function required by Smart Grids. Today increasing amount of smart metering data is available enabling the development of enhanced data-driven models for short-term load forecasting. Until now, a plethora of models have been developed ranging from simple linear regression models to more advanced models such as (artificial) neural networks (NNs) and support vector machines (SVMs). Despite the relatively high accuracy obtained, the acceptance of purely data-driven models such as NN models is still remained limited due to their complexity and nontransparent nature. Therefore it is important to develop optimization schemes, which can be used to facilitate the selection of appropriate model structure resulting good forecasting accuracy with low complexity. This study presents an optimization scheme based on multi-objective genetic algorithm (GA) for designing data-driven models for short-term forecasting of electric loads. The optimization scheme is demonstrated for designing the conventional NN/MLP model using real smart metering data and weather measurements. The optimal NN model structures are identified and analyzed in terms of model complexity and forecasting accuracy.

A data mining approach for producing small area statistics-based load profiles for distribution network planning

The recent European Union and national level initiatives such as INSPIRE and PSI have increased the availability of public sector data, which provides interesting new opportunities to support decision making in electricity distribution network planning. With big amounts of available data, data mining methods can be utilised to produce improved spatial load models. We propose a data mining approach, which uses the Self-organizing map for producing representative small area level load profiles based on building characteristics, demographics and automated meter reading data. Furthermore, the k-nearest neighbour algorithm and a genetic algorithm based feature selection are used in order to find a parsimonious set of features that can be used in selecting proper load profile. As the load profiles are based on area level statistics, they can be used to estimate the future loads in different scenarios regarding changes in population and building stock, which is particularly advantageous in distribution network planning.

Data-driven method for providing feedback to households on electricity consumption

The building sector is a major energy consumer and CO2 emitter, being responsible for approximately 40% of the total consumption in the EU. Active demand side participation of electricity customers is seen as crucial in the management and reduction of the building sectors CO2 emissions. However, todays electricity markets are often lacking strong incentives for active demand side participation. Understandable customer specific comparison information and easy-to-use energy displays can be used to influence customer behaviour and encourage customer participation. This paper presents a data-driven method for producing household level comparison information, based on hourly interval smart meter data and additional household information. Firstly, the customers are segmented by the heating system and the type of housing, followed by weighted clustering that is used to refine the comparison group. In the weighted clustering, normalized load profiles together with properties of the dwelling and the residents are considered, and weights are assigned to the properties according to how much they contribute to the electricity consumption. In this paper, the initial experimental results are presented and discussed, and future development ideas are laid out. The method is under development and testing as a part of the Finnish SGEM-project.

Computer-assisted image analysis of histopathological breast cancer images using step-DTOCS

In this paper, we address the epidemiology and morphology questions of breast cancer with special focus on different cell features created by lesions. In addition, we provide an insight into feature extraction and classification schemes in the image analysis pipeline. Based on our conducted research work, a novel feature extraction approach, a modification of Distance Transform on Curved Space (DTOCS), is proposed for analysis and classification of breast cancer images. The first experimental results suggest that the Step-DTOCS-based MLP-network is capable of discriminating different cell structures in a respectable way. The obtained results are presented and analyzed, and further research ideas are discussed.

Computer-aided breast cancer histopathological diagnosis: Comparative analysis of three DTOCS-based features: SW-DTOCS, SW-WDTOCS and SW-3-4-DTOCS

In this paper, the performance of three Distance Transform on Curved Space-based features derived from digital H&E stained oncopathological images used in breast cancer pattern recognition scheme are compared. The three features utilized are SW-DTOCS, SW-WDTOCS and SW-3-4-DTOCS with three different sliding window (SW) sizes. The results imply that the three distance transform features yield slightly different classification performance. In addition, the issue of computer-aided breast cancer histopathological diagnosis is discussed.