Electrical energy has to be generated whenever there is a demand for it. It is therefore imperative for the electrical power utilities that the load on their systems should be estimated in advance. For adequate electricity to be supplied to the customers, their load demand must be known. This estimation of load in advance is commonly called load forecasting. Due to a centralized power system and the continuous varying nature of the load, it is very difficult to balance demand and generation at all time [1]. Alternating current (AC) electricity cannot be stored and the science behind generation, transmission, and distribution of electricity requires that a stable or constant equilibrium be maintained between supply and demand in real-time. This can only be achieved by making reliable and accurate planning actions. If the system load forecast is overestimated, the system may overcommit the generation of power which will inadvertently lead to costly operation of the power system.
Zimbabwe's electricity generation is handled by Zimbabwe Power Company (ZPC) which is a subsidiary of Zimbabwe Electricity Supply Authority (ZESA). The utility has an installed capacity of about 2900 MW from two major sources for domestic electricity supply: which constitutes 58% thermal (coal-fired) and 37% hydro sources. In rural parts of the country, 80–90% of the people depend on wood fuel and kerosene for cooking and lighting [2]
Only 418 MW of the installed capacity was available on 21 January 2020 [3]. The difference is accounted for by age of the thermal plants and low water levels at Hydro-plants. Official figures from the Zimbabwe Power Company show that the country is currently generating an average of 418 MW from its five power stations against a daily peak demand of about 2100MW [3]. This means that ZESA now relies on regional imports to offset the deficit. Most of the units are now old and inefficient. For instance, the Hwange power station has an installed capacity of 920MW and averages 400–550 MW. The old coal-fired power stations in Harare, Bulawayo, and Munyati have a total installed capacity of 370 MW but most of the time they are not generating. The need for alternative energy supplies to meet the energy supply deficit cannot be overemphasized. To meet the electricity supply deficit, ZESA has to import expensive power from its neighbours, mostly South Africa (Eskom) and Mozambique (Cahora Basa). The difference in demand is accounted for by under-budget generation and unplanned generator outages among other infrastructural reasons. The electricity supplies cannot meet demand as the electricity grid is in a poor state due to inadequate investment in the sector, leading to erratic supplies and load shedding. The danger with importing energy from neighbouring countries is drawing more than allowable inadvertent power of 95 MW from the Southern African Power Pool (SAPP) interconnected power system which attracts stiff penalties. ZESA has drawn more than the allowable advertent power of 95 MW from the SAPP system many times. As a result, this has reduced the revenue to import energy. Moreover, power outages continue to affect the economic performance of industries and services. Furthermore, poor forecasting methods have led to the distortion of the nominal frequency of 50Hz across the interconnected grid, and the power exchange threshold of 95MW has been violated several times while trying to restore the nominal frequency giving rise to the penalties. Proper techniques to forecast the load are hence imperative to eliminate these stiff penalties. This will enable power utilities to make economically viable decisions regarding future generation, transmission, and distribution. Also, the utilities will promote maximum utilization of power generation plants and avoid under-generation or over-generation and hence guarantee economic dispatch.
According to [4], there are 3 types of load forecasting which are short term load forecast which varies from 1 hour to one week and medium-term load forecast which varies from a week to usually a year, and lastly long-term load forecast which is longer than a year. Consideration of various factors is the prerequisite for accurate forecasting of the load. Whilst trying to keep the system reliability in reasonable tolerances, the forecasted system load must be met at the lowest possible cost which is a key result area in power system operation [5]. According to a review by [6], more than 113 different case studies across 41 academic papers have been used for the selection process. Factors such as time frame, inputs, outputs scale, data sample size, error type, and value were considered as criteria for the comparison. The review shows that despite the correlative nature of all reviewed models, the regression/multiple models are still very efficient and default to solving long and very long-term forecasting. [7] alludes that load forecasting at an individual household level is a challenging task that requires extracting load data directly and knowledge of the individual load profiles which are influenced by various factors. These factors include device operational characteristics, user behaviour, time of the day, and so on.
According to a review by [8], various methods can be used to solve SLTF which include Artificial intelligence (AI), Statistical techniques (non-linear regression, regression trees), and knowledge-based expert systems (fuzzy expert systems). Time-series methods treat the load pattern as a time-series signal with known seasonal weekly and daily periodicities. These periodicities give a rough prediction of the load at the given season, day of the week, and time of the day. The difference between the prediction and the actual load can be considered as a random signal [9]. A statistical approach to load forecasting using a regression model was done for the Zimbabwe Electricity Supply Authority (ZESA). The study aimed to construct an effective simplified econometric model that can be used to forecast the peak demand for electricity for Zimbabwe from 2011 to 2015. The use of three different software environments, MATLAB, SPSS, and Excel confirmed that they all have sufficient statistical capability to carry out reliable modelling. Results from MATLAB were most preferable because of their simplicity. The correlations investigation revealed that there was high collinearity with a Pearson correlation significant at the 0.01 level (2-tailed for: Peak demand with population and GDP per capita, maximum temperature, minimum temperature, and temperature range. The set of customized growth curves that align to different economic sectors and the constant percentage growth algorithms are used to align the electric forecast to the econometrics study [10]. However, according to [11] econometric methods have been long outmatched by other techniques due to lack of forecasting accuracy compared to forecast from simple mechanical schemes and autoregressions. Therefore, they are not advocated for in this application. Spatial forecasting has been also applied by Eskom, a South African power utility but due to this forecasting method, South Africa experienced serious power shortages in 2007. According to a review by [6], more than 113 different case studies across 41 academic papers have been used for the selection process. Factors such as time frame, inputs, outputs scale, data sample size, error type, and value were considered as criteria for the comparison. The review shows that despite the correlative nature of all reviewed models, the regression/multiple models are still very efficient and default to solving long and very long-term forecasting. Regional load forecasting involves predicting the amount of electricity that should be generated to supply specific kinds of consumers over a specific period and location.
The national grid which is based in the United Kingdom is collaborating with DeepMind®, a google owned Artificial Intelligence (AI) team. Deep mind is working on how to predict the power supply and demand peaks using statistics generated from the smart meters and weather stations within the UK. By using deep learning technology and machinery, demand and supply could be predicted and controlled in real-time as a result of load dispatch is optimized and operation costs are lowered [12]. According to [13] an ANN can be defined as a highly connected array of elementary processors called neurons. Artificial Neural Networks (ANNs) refer to a class of models inspired by the biological nervous system. The models are composed of many computing elements, usually denoted neurons; each neuron has several inputs and one output [14]. The figure below shows a single-layer perceptron.
In the above Fig. 1 [15], for one single observation, x0, x1, x2, x3...xn represents various inputs(independent variables) to the network. Each of these inputs is multiplied by a connection weight or synapse. The weights are represented as w0, w1, w2, w3….wn. Weight shows the strength of a particular node, b is a bias value. Mathematically, the sum of the products of weights and observations gives,
Adding bias b, to Eq. 1, gives:
According to [16], The most important unit in neural network structure is their net inputs by using a scalar-to-scalar function called "the activation function or threshold function or transfer function", output a result value called the "unit's activation". An activation function for limiting the amplitude of the output of a neuron. There are several activation functions which include Sigmoid Function, hyperbolic function, rectified linear unit activation function, and so on [17]. A Multi-Layer Perceptron (MLP) contains one or more hidden layers (apart from one input and one output layer), as shown in Fig. 2 below [18]. While a single-layer perceptron can only learn linear functions, a multi-layer perceptron can also learn non-linear functions, which makes it suitable for learning nonlinear, complex relationships hence their application in load forecasting.
In neural networks, to feed-forward is to give a certain input to the neural network, this can be historical load. The network will calculate the output by propagating the input signal through its layers. In other words, the output form one layer becomes the input to the next one, where the output from the last one is the final answer [19]. It has also been applied to load forecasting before.