Comparative analysis and assessment of ANFIS-based domestic lighting profile modelling

Highlights

• Addresses and solve non-linear issues, ambiguity and randomness of data for adept estimation and prediction of lighting load profiles.

• Better correlation in comparison with other research studies and models was also deduced.

• Can be applied in various settings such as non-availability of housing database, layout details, difficult terrain, etc.

• Being an income model, it is able to effectively to extract, interpret and infuse social norms and values, activities etc within a group on lighting usage modelling.

• ANFIS-based learning capability and aptitude to reproduce gives it an edge in prediction accuracy.

Abstract

A good number of stimulations tools reproduce the deterministic physical behaviour of buildings especially in lighting with repeated standard patterns of occupant presence without replicating the dynamic occupancy and activities associated with such environment. This thereby contributes to peak energy/demand crisis being experienced in countries and over estimation of energy savings associated with energy (lighting) efficient projects that have been embarked upon by utilities or government. This research work involves the comparative and performance assessment studies of an ANFIS-based model that is capable of addressing and solving non-linear issues, ambiguity and randomness of data to ensure adept estimation and prediction of lighting load profiles. The proposed technique is based on learning and adaptation of the variables associated with lighting usage. Two different investigative approaches were applied in relation to the ANFIS-based model for domestic lighting profile development. Validation process was carried out in terms of the model profiles to ascertain the performance of the methodology. Good correlation and coefficient of determination in comparison with the actual output; better correlation in comparison with other research studies and models was also deduced. This technique is expected to assist utilities, energy and measurement and verification practitioners as well as contribute to lighting load profile modelling.

Introduction

Almost every aspect of modern life is based around electricity. Countries are facing challenges as a result of increasing demand for electricity. Eskom, which accounts for 95% of electricity supply in South Africa, has experienced many difficulties in the past ten years, failing to reach the acceptable energy reserve margin of 15 to 20%. This imperatively may lead to load shedding in winter, even though various demand side management (DSM) strategies are being implemented [1].

DSM is the implementation of policies and measures, which serve to control, influence and generally reduce electricity demand. DSM aims to reduce energy demand and consumption, while preserving an acceptable level of electricity generation without compromising service delivery and comfort of the consumers. The term “DSM” according to Grover and Pretorius [2] can be summarized as “the planning and implementation of utilities activities designed to influence the customer to use electricity in ways that will produced desires changes to the utility's load shape”. In South Africa primary activities such as energy efficiency and energy conservation, load control (an amalgamation of load shedding, valley filling and peak clipping strategies)—specifically targeting load peak capacity are being implemented by Eskom [2]. Eskom has embarked on various DSM strategies and initiatives to reach its target in order to avoid load shedding, however the improper quantification of energy usage or savings from these strategies have imparted on their effectiveness [1]. Proficient estimation of load profiles (lighting) is crucial for demand side management and implementation of proposed energy efficient project across board. One of the measures for conservation, reduction and management of electricity is lighting. It is an essential part of human daily activities especially in households. The impact of major variables on lighting energy consumptions pattern in buildings could assist in the load profile development and evaluation of demand management initiatives highlighted above.

Modelling techniques are quite imperative in the provision of good estimation and forecast in daily optimal use and management of resources to ensure consumers (customer) satisfaction. Analytical, empirical and modelling methods are existing techniques available for prediction of energy use or lighting load [3], [4] Most of these methods do not rely on occupant behaviour from survey data/long-term observations. Others are mostly based on conventional mathematical tools, which are not well suited for ill-defined or uncertain (behavioural) system. Because of such complexities, good lighting demand or energy prediction accuracy are rarely found. Identification of major determinants of energy consumption in buildings with good understanding of such determinants impact on energy consumption patterns will assist in energy performance and reduction of greenhouse gases [5]. Some of these determinants are climate (e.g. outdoor air temperature, solar radiation, wind velocity etc), building occupants’ behaviour and activities, social and economic factors, building related characteristics (type, area and orientation) etc [5]. Other simulation tools based behaviour of residential occupants on assumptions thereby into resulting prediction models thus providing a poor tool to evaluate and predict demand initiatives outcomes [6]. Findings also suggests that interaction of an individual are difficult to predict, however behavioural trends and patterns for building occupants can be extracted from long-term historical data [7].

Occupant behaviour (occupancy) is an important aspect in the usage of energy. Discrepancy in energy consumption among buildings of similar constructions can be attributed to occupancy patterns and occupant behaviours, hence the need for building occupant behaviours when designing energy programs or models however there is difficulty associated with doing so [8], [9]. Studies conducted associated with occupant behaviour in relation to occupant presence and activities include stochastic and mathematically derived approaches. Some of the studies had shortcomings such as inability to capture a number of the behavioural patterns [10], [11], [12], [13], [14].

Numerous other studies have shown that energy consumptions in residential household are income correlated. One of the papers presented information's that income distribution play a vital role in the acquisition of energy related goods in rural China residential households [15]. According to the authors, this factor is generally omitted from forecasting models. Studies have also shown that usage of different appliances and electric lighting being used by the occupants is strongly influenced by the outdoor global irradiance thereby impacting on customers load profiles [16], [17]. Although there exists a platform for the integration of occupant models i.e. occupants’ presence and their effect on the building, there is to date no complete and interlinked set of models considering all aspects of occupant behaviour. In addition, the most advanced published models for occupant presence still neglected its time-dependence over a day/year [18].

Occupant's actions are strongly dependent on the occupant's behaviours, responsibilities and other factors [6]. Mostly this is necessitated by the impact of the environment. The complexity, non-linearity and characteristics associated with lighting usage have created a need for more research interest. There is the need to evolve a modelling technique that can generate lighting load profile with such characterization and factors like occupant presence, income (comfort level), outdoor environment factor (natural light or solar irradiance level) etc. Research has also shown that government and organizations have looked to lighting as swift means of achieving energy reduction because of electricity demand outstripping generation in developing countries [19]. This study carried out a comparative and performance assessment of ANFIS-based model for lighting load profile development.

The proposed investigation methodology is known as the adaptive neural fuzzy inference system modelling (ANFIS). Features of ANFIS include the ability to learn, more translucent, organizes network structure itself and adapts the parameters of the fuzzy system to solve engineering problems [20], [21]. ANFIS can simply be defined as a set of fuzzy ‘if–then’ rules with appropriate membership functions (MF) to generate the stipulated input–output pairs in the solution of uncertain and ill-defined systems. In determination of the membership functions, learning provision of the ANFIS and construction of the rules, backward propagation algorithms and hybrid-learning algorithms methods are applied. In broad ANFIS system has input layer, output layer, and hidden layers that represent membership function and fuzzy rules.

General three rule being: if x is A_i, y is B_i and z is C_i, then

where x, y and z represents inputs which are fuzzy sets A_i, B_i, and C_i representing natural lighting, occupancy and income. p, q, and z are the design parameters that will be determined during the training process. For instance the input nodes i.e. first layer (i.e. Q_i¹) has the following inputs with a node function output [3], [4], [22].

$$Q_i^1={\mu }_{A_i}\left(x\right)\text{for}i=\mathrm{1,2,3};Q_i^1{=}_{B_i}\left(y\right)\text{for}i=\mathrm{4,5,6};Q_i^1{=}_{C_{i-3}}\left(z\right)\text{for}i=\mathrm{7,8,9}$$

where x is the input to node _i, and A_i is the linguistic label (high, middle, and low) associated with this node function. ${\mu }_{A_i}$, ${\mu }_{B_i}$ and ${\mu }_{C_i}$ are the appropriate membership functions of A_iB_iC_i fuzzy sets. Nodes in this layer represent the MF associated with each linguistic term of input variables. A trapezoidal membership curve was chosen for this investigation due to human behavioural inclination (occupancy and natural lighting patterns) and historical lighting load metering profile for this investigation. The trapezoidal membership is a function of a vector x, and depends on four parameters a, b, c, and d, as given by (2). a and d parameters locate the “feet” or base of the trapezoidal and b and c the parameters locate the “shoulder”.

$$f\left(x;a_i,b_i,c_i,d_i\right)=\max \left(\min \left(\frac{x-a_i}{b_i-a_i},1\frac{d_i-x}{d_i-c_i}\right),0\right)$$

The other layers include: second layer outputs the firing strength w_i, third layer calculates the ratio of ith rule's firing strength to the sum of all rule's firing strengths while the fourth layer computes the ration of the firing strength and consequent parameters. The output nodes (fifth layer (Q_i⁵) compute the overall output as the summation of all incoming signals. The defuzzification process is used to transform each fuzzy rule which results into a crisp output as shown below. N represents the number of rules in the investigation [22].

$$Q_i^5=f\left(\text{out}\right)=\frac{{\sum }_{i=1}^N{w}_i\times f_i}{{\sum }_{i=1}^N{w}_i}$$

For the fuzzy light usage output design, input variables natural light level, effective occupancy (Active/Awake occupancy) and; income of a household were applied [22]. The process design strategy shown in Fig. 1 was applied in the model development strategy.