Application of Neural Networks for the Prediction of the Energy Consumption in a Supermarket

It has been shown by previous researchers that Artificial Neural Networks (ANNs) not only be used to predict energy more reliably than traditional simulation models and regression techniques but can also from the basis for a predictive controller of thermal systems such as HVAC equipment. This work is directed towards the identification of the important inputs (independent variables) to facilitate on-line prediction and thereby implement refrigeration and HVAC system diagnostics, process control, optimisation and energy management in retail food stores. This paper presents preliminary results on the prediction of electricity consumption with different independent input variables in a supermarket. The paper also compares the prediction performance of neural networks with the more traditional multiple regression techniques. INTRODUCTION Most Supermarkets in the UK qualify to join the competitive power market and thus they purchase power from a pool established by the Power generating companies on a half hourly basis. The half hourly rate paid by the supermarket owner to the supplier is to a large extent depended on his ability to predict accurately his maximum halfhourly demand and the competitive rates offered by various suppliers. If the actual consumption exceeds the predicted value, the purchaser is penalised for the extra supply needed by paying at a higher than the negotiated rate. The ability, therefore, to predict the power consumption every half hour as accurately as possible will facilitate negotiations on electricity tariffs with the suppliers and will also enable the control of maximum demand by shifting some of the load to periods of reduced demand. To date, neural networks have been applied successfully to a number of engineering problems. Several researchers have demonstrated that they can be more reliable at predicting energy consumption in a building than other traditional statistical approach [1,2,3,4] because of their ability to model non-linear patterns. The neural network learns the main characteristics of a system through an iterative training process. It can also automatically update its learned knowledge on-line over time. This automatic learning facility makes a neural network based system inherently adaptive. Furthermore, its predictive capability can be used to optimise the operations of the refrigeration, heating and ventilation plant in a building or supermarket, and thus spread the demand of power over the day, reducing maximum demand charges. This paper address the performance of a neural network in the prediction of electricity demand in a supermarket. It evaluates the capability of a neural network to forecast the overall power consumption of the store every half hour with respect to time of day and environmental conditions. A comparison of the prediction performance of the network against more traditional statistical approaches is also presented. ARTIFICIAL NEURAL NETWORKS A neural network is a non-linear mapping of the space between an input data set and an output data set and consists of three parts an input vector (independent variables), an output vector (dependent variables), and an algorithm that maps the input space to the output space. One or more hidden layers connect the external layers by a set of “weights”, expressed as two-dimensional matrices, W. In a feed-forward neural network, the value of each node in a particular hidden layer is the result of a nonlinear transfer function whose argument is the weighted sum over all the nodes in the previous layer plus a constant bias B. A variety of training algorithms are available but in general, to train a network, one begins with a set of training data consisting of the input vector, and corresponding target vector, T X 0 m. The internal weights are adjusted until the sum of differences between the neural net outputs Ym and the corresponding target Tm is minimised to a predetermined level for all the training data. A neural network with zero hidden layers is a linear expansion and a network with one hidden layer and a single output can be represented by[5]: 2 B + (1) ( ) ( ) Y W i F W i j X j m i N m j N = = = ∑ ∑ 2 1 1 0 1 1 1 0 ( ) . In the above equation, N1 is the number of nodes in the hidden layer and N0 is the number of independent variables. A sigmoidal function is usually used for the transfer function F as it enables a finite number of nodes in the single hidden layer to uniformly approximate any continuous function. Training of neural networks is frequently performed using a backpropagation algorithm. This algorithm iteratively adjusts the weights to reduce the error between the actual and desired outputs of the network. Detailed descriptions of different network configurations and training techniques are given by Rumelhart and McClelland[6] amongst many others. One facet of neural networks is that a statistical understanding of the relationships between the independent and the dependent variables is not needed. Continuous analysis of the independent variables can lead to well chosen network inputs. 3 TRAINING AND TESTING OF NEURAL NETWORKS Supermarkets are one of the largest single end users of electricity with refrigeration systems accounting for more than 50% of the electricity used. About 25% is accounted for by the HVAC equipment and lighting while the other utilities account for the remainder. The energy consumption of supermarket refrigeration systems is a function of a number of variables which include the building fabric, the ambient conditions (temperature and humidity) and the internal environment. In the UK, it is a common practice for the refrigeration and HVAC system to be part of an integrated design taking advantage of the rejected heat from the refrigeration packs to provide heating in the store. In trying to minimise energy consumption, therefore, the various energy consuming subsystems cannot be viewed in isolation but their interactions should be considered as well as their influence on the sales revenue and profitability of the store. Energy consumption can be minimised only through better understanding of the consumption patterns and better control of the major energy consuming equipment in response to external and internal environmental conditions. Computer based monitoring and control systems provide the opportunity to characterise the various energy consuming processes in the store and relate the consumption patterns to fuel pricing and tariff structures. These systems can further be developed to incorporate advanced control techniques to minimise maximum electricity demand, energy consumption and fuel costs. To this end, we propose the use of Artificial Neural Networks because unlike traditional system modelling techniques ANNs are not system specific and can be easily adapted to different building types, HVAC systems and refrigeration equipment. The current investigations are based on a Safeway supermarket situated in Airdrie, Scotland. This store has been equipped with a central monitoring and control system which monitors the temperatures of the display cases in the store and controls the refrigeration packs. For the purpose of the project the system has been extended to incorporate a number of additional variables which include: Temperature and relative humidity in the store, External air temperature and humidity, Total electrical power consumption of the store, Electrical power consumption of the refrigeration packs, Gas Consumption and Underfloor heating flow and return temperatures. Simple three layered feed-forward neural networks were trained using the actual measured data collected from the store. The networks varied in terms of the number of input variables, i.e. input nodes, n. The number of nodes of the hidden layer varied as a function of the input nodes as (2n + 1). The standard back-propagation algorithm was employed to train all the networks. Back-propagation is an integrative training algorithm designed to minimise the mean square error between the output of the network and the actual value.