Short term load forecasting using a synchronously operated recurrent neural network

A keypoint of the control of a power system is the forecast of the short term load. This paper presents a dynamic model for short-term load forecasting (STLF) which uses a recurrent neural network. This network can be used to build empirical models for the load of a dynamic system. We investigate this problem applying a basic neural network with feedback connections which is unfolded in time and becomes a general feedforward network with weights sharing. The main advantage of this model consists in unfolding the network in time, which becomes a non fully connected feedforward network and facilitates the training stage. At the same time our model provides a one day ahead prediction. Introduction The close tracking of the electric load by the dispatcher is a basic requirement in operations of power systems. Load management is a procedure used from longterm planning to daily operation of a power system. The main aim of load management is to avoid the spreading of an emergency in the situations of unbalance between generation and load in the system and the overload of equipment in a part of the system. The optimization of the load curve may be reached by a simple differentiation of tariffs to encourage the customer to use less electric energy at the peak time of the system. The other way is to give the customer an economic advantage for allowing the load dispatcher the reduction of the load. For avoiding a system collapse, the load management must be done by automatic equipment or through direct switching by the dispatcher. Hence, suitable strategies are necessary for load management and optimum operation of electric power system. For this purpose, shortterm load forecasting (STLF) has to be carried out as accurately as possible. The modelling of the power system load and its prediction is essential for the economic and reliable daily and weekly operations scheduling. In addition, the load model and forecast are essential information for security analysis in both the real-time and study modes. Up to now, various forecasting techniques have been tested to solve power system load forecasting, such as multiple variable regression, weather sensitivity analysis, state space methods, time series methods, expert system methods, artificial neural networks, and so on [1], [2], [3], [4], [5] In this paper the construction of a dynamic model for STLF using a recurrent neural network is presented. In nature, complex dynamic patterns, like the system load profile, are often the result of relatively simple underlying generating mechanism. The recurrent neural networks can be used to construct empirical models for the load as a dynamic system [6]. This paper investigates a basic neural network with feedback connections, which is unfolded in time and becomes a general feedforward network with weights sharing. The basic backpropagation is a method for calculating all the derivatives of a single target quantity with respect to a large set of inputs. Backpropagation Through Time algorithm, studied in this paper for an actual problem of STLF, extends the basic backpropagation so that it applies to dynamic systems. This allows one to calculate the derivatives needed when optimizing a neural network with memory. The main advantage of our model consist in unfolding in time the network, which becomes a non fully connected feedforward network and facilitates the training stage. In the same time our model provides the one day ahead prediction. The study has been based on three years hourly load data provided by an Italian utility. It can also consider the weather effect on power demand. Dynamic neural model The system load profile can be modelled as the output of some dynamic system, influenced by weather, time and other environmental variables. A neural network with feedback can simulate a discrete time dynamic system. The general feedforward topologies with weights sharing can represent feedback connections by unfolding in time the basic network. Of course by unfolding in time the network, ( ) ( ) ( ) ( ) ( ) ( ) ( )