A Comparative Evaluation of Monthly Electricity: Consumption Forecasting in Sri Lanka

Do you need this or any other assignment done for you from scratch?

We have qualified writers to help you.

We assure you a quality paper that is 100% free from plagiarism and AI.

You can choose either format of your choice ( Apa, Mla, Havard, Chicago, or any other)

NB: We do not resell your papers. Upon ordering, we do an original paper exclusively for you.

NB: All your data is kept safe from the public.

Click Here To Order Now!

A Comparative Evaluation of Monthly Electricity: Consumption Forecasting in Sri Lanka

Abstract

Sri Lanka as a developing country, over 98% households have been electrified and it is crucial to plan for future electricity demand in order to match the demand with supply. This study aims at forecasting monthly electricity consumption in Sri Lanka and explores the weather influence on that electricity consumption. Due to higher living standards, weather has a considerable impact on the short-term electricity demand due to the use of fans, air conditioners and refrigerators. There are three main weather-related factors that affect demand they are Rainfall, Humidity and Temperature. In this study, four forecasting approaches (Classical Decomposition Method, Exponential Smoothing Method, Probabilistic Modeling in Time Series, Autoregressive Distributed Lag (ARDL) model Approach) were employed in forecasting monthly electricity consumption in Sri Lanka. Additionally, Artificial Neural Network (ANN) was used in order to investigate the applicability on forecasting monthly electricity in Sri Lanka as an advanced method. Arc-map software was used for spatial computations and twenty meteorology stations were considered to spatially interpolate the weather data by using the Inverse Distance Weighted (IDW) interpolation method. Under this study, it was revealed that ARDL and ANN model approaches which the weather influence was incorporated perform better in monthly electricity consumption forecasting.

Keywords: Electricity consumption forecasting, Weather impact, Inverse distance weighted interpolation, Autoregressive distributed lag, Artificial neural network.

Introduction

Electricity that is delivered to end-users like households and businesses is the result of a complicated operational process starting from Generation, Transmission, Distribution and real time supply-demand balancing. Ceylon Electricity Board (CEB) being the only generation and transmission utility in Sri Lanka is responsible for this total cycle and this operation is controlled centrally by the System Control Centre (SCC) which is the nerve center of operational activities of the CEB. As electrical power in Alternating Current form (or AC in short) cannot be stored in real-time, the SCC need to match the demand for electricity in the country at any given time with exact amount of generation in real-time. SCC needs to meticulously plan this real-time supply-demand balancing act and this type of planning is called “operations planning”. Operation planners need to plan for time steps from next week to several months ahead. One of the fundamental prerequisites of Operations Planning is short-term demand forecasting.

The main objective of this study is identifying the correlation between past actual monthly electricity demand against the historical three weather parameters (temperature, rainfall and humidity) and to present System Operators with a better model to do short term demand forecasting by inputting forecasted weather parameters which are available with forecasting agencies for few months ahead. Second objective of this study is to identify a better forecasting method in order to forecast monthly electricity consumption in Sri Lanka by comparing the forecasts done based on weather influence with other potential forecasting approaches.

The availability of valid, efficient and effective forecasting models for electricity loads is successful way of reducing the need for new generation capacity, scheduling resources, and by improving efficiency in electricity end use. In a broader economic sense, such models help in achieving a more efficient allocation of resources. Unlike long term expansion planning where the demand growth over 20 year time horizon is driven by external factors such as economic growth, population, tariff, etc, short term forecasting is primarily driven by weather and short term social factors such as holidays and festivals[1]. Due to higher living standards, weather has a considerable impact on the short-term electricity demand due to the use of fans, air conditioners and refrigerators [1]. There are three main weather related factors that affect demand. They are Rainfall, Humidity and Temperature [2]. At present, System operators (System Control Engineers) do the short term demand forecasting using time trend analysis and using their experience. Within a day, forecasting (which is done for shorter time steps of hours) is primarily done using historical hourly demand curves whereas weekly forecasting is done considering past data (in daily time steps). However, in system control center monthly forecasts (in monthly time steps) are done only based on time trend analysis. Therefore, in this study, it is anticipated to consider the weather variation in the country in forecasting the monthly electricity consumption.

Methodology

There are 62 CEB areas as covering the country and for these entire areas monthly tariff wise consumption and consumer accounts data were obtained for the period of January 2006 to June 2017 from Information Technology Branch, CEB. Monthly minimum and maximum air temperatures, minimum and maximum relative humidity and rainfall data along with latitude and longitude were collected from the Meteorology Department and Department of Census and Statistics for 20 meteorology stations as these stations are sufficient to cover the whole country. Under this study, starting from the simple forecasting methods up to complex methods various approaches were employed in forecasting monthly electricity consumption in Sri Lanka. They are,

Classical Approach of Time Series Analysis
Exponential Smoothing Method
Probabilistic or Stochastic Modeling Approach
Autoregressive Distributed Lag (ARDL) modeling approach
Artificial Neural Network

The original series of total monthly electricity consumption which was prepared by adding CEB area wise data, was divided into two parts as training and test set and the period of January 2005 to December 2015 was used to train the model for all forecasting methods while keeping the period of January 2016 to June 2017 (18 months) as the test set.

A time series is a collection of observations of some variable made sequentially in time, in which the time intervals are in regular intervals. Classical approach identifies four components (trend, cyclic fluctuations, seasonal variations and irregular variations) which affect time series values. Exponential Smoothing method is a very popular method in producing a smoothed time series. In single moving averages the past observations are weighted equally while Exponential Smoothing assigns exponentially decreasing weights as the observation get older. That means recent observations are given relatively more weight than the older observations when forecasting is done. The goal of the probability models for time series is to characterize random component of the series. The foundation of time series analysis with probability models is stationarity.

In order to deal with ARDL models and ANN models, weather variables were used as inputs. There were number of missing values in weather variables and to overcome this issue, missing values of weather parameters were interpolated through a computer program written in R (R version 3.4.1.). “imputeTS” is the only package on CRAN that is solely dedicated to univariate time series imputation and it includes multiple algorithms[3]. Among them most popular four algorithms were used and they are, Interpolation, Kalman, Seadec and Seasplit [4]. Performance of these algorithms can vary according to the structure of the series. Figure 1 contrasts the values imputed by these algorithms for a series of missing values of minimum temperature with actuals. To analyze the efficiency of the mentioned algorithms, its performance is compared with the root mean squared error (RMSE) which was chosen as a performance measure.

Figure 1: Actual values and imputed missing values by the algorithms

Model Building for Sub Regions in the Country

Sri Lanka as a tropical country has been traditionally divided into main climatic zones, namely Wet, Dry and Intermediate zones. Thus the weather variation is not uniform over the country. Hence unlike the first three forecasting approaches, the total monthly electricity consumption of the country cannot be taken as the response variable when influence of weather parameters on the consumption is in concern for ARDL and ANN approaches. Therefore, the need of dividing the country into sub regions such that each sub region possesses similar weather conditions was aroused.

As the simplest way, this categorization could not be done according to the climatic zones in the country because the electricity consumption was measured based on CEB areas. With this limitation the optimal way in order to get rid of this obstacle was matching the CEB areas into districts according to their geography boundaries. Then assigned the weather parameter values which were imputed for district centroids as each districts’ representative weather parameter values. Thus there are 20 districts and 3 CEB areas which were taken for region wise monthly electricity consumption forecasting.

In order to modeling the district wise and CEB area wise monthly electricity consumption, the meteorological parameters should be fair enough to represent the particular district’s or CEB area’s weather condition. Hence representative weather parameters have to be imputed for a specific location of each and every district and CEB area. Therefore as representative locations, centroids of the districts in Sri Lanka were found using Arc Map 10.3. The conversion of longitude and latitude into SLD99 systems was done for the 20 meteorology stations for the further applications in Inverse Distance Weighted (IDW) Interpolation and in order to extract the imputed values for centroids of districts. This was done using Arc Map 10.3. Due to sparseness of the stations, advanced methods such as “Kriging”, which accounts for the spatial correlation, is inappropriate. Hence distance based IDW interpolation was employed in estimating the weather parameters at district centroids. IDW is a spatially weighted average of the sample values within a search neighbourhood. In IDW interpolation method, the sample points are weighted according to distance and the influence of one point relative to the decreases with distance from unknown point which we are interested.

Figure 2: IDW Interpolation for average relative humidity for January 2006

Figure 2 represents the interpolation for average relative humidity for the month January 2006 and thus for all weather variables IDW interpolations were applied for all months (138 months). After identifying the relevant grid points, for each variable in each month the imputed values were gathered and those values are illustrated from Figure 3.

Figure 3: Time Series Plot of Interpolated Minimum Relative Humidity by districts

ARDLs are standard least squares regressions which include lags of both the dependent variable and explanatory variables as regressors. Then ARDL models were fitted to each region by incorporating the weather variables. In order to estimate the total electricity consumption in the country, ARDL models were built for 23 regions and by adding up all the forecasts given by each model, total forecasted consumption was obtained. In ARDL forecasts, there were two options as dynamic and static and compared to the dynamic forecasts, static forecasts which make the next step forecast using actuals was obviously better than dynamic forecasts where previous forecasted values taken in next step forecast.

ANNs are able to learn and model non-linear complex relationships. Unlike other prediction methods, ANNs do not impose any restrictions on the input variables[6]. Additionally, from some studies it has shown that ANNs are better even in modeling data with high volatility and non-constant variance. ANN was built to determine the performance of other forecasting approaches with its performance [7].

For the first three approaches total monthly electricity consumption in the country was considered as a univariate time series and forecasting was done while in ARDL model, since weather parameters were used, the district-wise and CEB area-wise consumption forecasting had to be done. Ultimately adding up all the forecasted values given by relevant ARDL models for each district and CEB areas, the total forecasted monthly consumption for Sri Lanka was calculated. Then the actual total monthly electricity consumption was compared with these first four methods. ANN approach was used for forecasting monthly electricity consumption in one region and it was compared with the relevant ARDL model for the same region in the purpose of comparison.

Results and Discussion

Under Classical approach and Exponential Smoothing method, additive models were the best model between those approaches’ Additive and Multiplicative models. Hence they were used for this comparison. Figure 4 graphically shows the forecasts by each four methods and the actual total monthly electricity consumption for the last 18 months which was undertaken as the forecasting period or test set.

Figure 4: Comparison of Forecasts of the Total Consumption with Considered Forecasting Models

Figure 5: Predicted vs. Actual of the Total Consumption for the Forecasting Models

According to the Figure 5, where the predicted vs. actuals by forecasting methods are presented in the same plot, it can be mentioned that the SARIMA model performs much inadequately compared to other three methods and the Additive model of Classical method and Additive model of Triple Exponential Smoothing models performs approximately in a similar way. Among the all four methods it was apparent that forecasts by ARDL approach is most probably performing better over other three methods.

Figure 6: Comparison of MAPE for the Forecasting Models

Accuracy measure MAPE for compared methods is graphically shown in Figure 6 and ARDL approach has the smallest value for MAPE. Hence it can be said that ARDL approach outperforms other three methods in forecasting. The performance of these forecasting methods can be listed in descending order according to MAPE as follows.

ARDL modeling approach
Exponential smoothing method
Classical approach
SARIMA modeling approach

As an overall, the weather variables used in this analysis are the imputed values by IDW interpolation, and therefore if it is required to forecast for a future month, then forecasts for weather variables provided by weather forecast agencies should be collected for these 20 meteorology stations and thereafter weather data should be imputed for centroids of districts by IDW interpolation. It can be presumed that if the actual weather data were available at district centroids ARDL and ANN models might be more preferable. The comparison between ARDL and ANN models was done using only one region. If it had been built 23 ANNs like in the case of ARDL modeling approach, via this generalization the performance of forecasting could be much better over all these forecasting methods. Thus it can be assumed that rather than a linear relationship there could be a non-linear relationship between the inputs and the response in this study. Therefore ANN can be suggested for further studies in forecasting monthly electricity consumption in Sri Lanka.

It was evident by the experience of system operators in SCC that the electricity consumption drops significantly in mercantile holidays. A dominant reason for this is that during the mercantile holidays industries are having a short break. However, in the ARDL models built for Trincomalee and Nuwara Eliya, number of mercantile holidays per month was significant but with smaller coefficients. A possible reason for this is that since both districts are known as tourist destinations due to migration during the holiday seasons, consumption of these areas could have increased.

This study was based on monthly electricity consumption and if it had been dealt with the daily consumption and weekly consumption, fluctuations of electricity consumption in mercantile holidays could be more sensitively captured. Although we have to deal with monthly data due to the limitations of obtaining data, weather influence on consumption also can be captured better if daily and weekly consumption were taken along with daily weather parameters.

It is a common fact that power failures are possible and due to that consumption may reduce, which has been neglected in this study. The LECO (Lanka Electricity Company Limited) sales are not included in the obtained data from IT branch of CEB and it was ignored under this study. The consumption data is updated to CEB data base through the meter reader. Meter reader visits and reports the consumption once a month and previous meter reading is subtracted from the current reading. That is how the monthly consumption is recorded as billing months and often meter readers visits happen in the middle of the month. Therefore the obtained monthly billing data might not be the corresponding accurate consumption for each month and it may include some portion of the actual consumption of previous month and the other portion from the actual consumption of current month. In future, CEB is going to establish smart meters with Subscriber Identification Module (SIM) and thereby the meter can be read remotely from CEB premises. With these data, forecasting can be developed further.

Conclusions

The main objective of this study was to find out the best forecasting model in order to forecast monthly electricity consumption in Sri Lanka and explore the influence of weather variables on the consumption behavior. Electricity consumption is highly correlated with the number of consumer accounts and weather variables. Correlation with weather variables depends on the monsoon season. Consumption of most of the districts had a moderate linear correlation with weather parameters. In the months of April and May, due to Sinhala and Hindu New Year festive seasons, the total electricity consumption was higher than that of in other months of the year.

Additive model performed better than multiplicative model in forecasting total monthly electricity consumption under Classical approach. In Exponential Smoothing method, the Holt-Winters’ Additive model was accepted as the best model compared to the Multiplicative model based on the measures of accuracy. model was the selected model from several candidate models in Stochastic approach. Since all the weather variables show a seasonal pattern, when missing values are distributed as a set of successive points, “Seadec” and “Seasplit” algorithms were always better in missing value imputation in weather variables. But in situations like one or two missing values have been distributed over the series, then any of the algorithms of missing value imputation was reasonable. Among all forecasting methods considered, ARDL model was the most appropriate model in forecasting the monthly electricity consumption. Also ANN models appear to perform even better than ARDL approach. Since both ARDL and ANN can be handled through other variables which are having influence on response variable, it can be concluded that rather than considering a single time series for the whole country it is better to deal as regression type models in forecasting monthly electricity consumption in Sri Lanka. Finally, it was found that weather variables have a considerable influence on the behavior of monthly electricity consumption in Sri Lanka.

References

Alex D Papalexopoulos and Timothy C Hesterberg. A regression-based approach to short-term system load forecasting. IEEE Transactions on Power Systems, 5(4):1535-1547, 1990.
James W Taylor and Roberto Buizza. Using weather ensemble predictions in electricity demand forecasting. International Journal of Forecasting, 19(1):57-70, 2003.
Steffen Moritz and Thomas Bartz-Beielstein. imputeTS: Time Series Missing Value Imputation in R. The R Journal, 9(1):207-218, 2017.
David S. Fung. Methods for the estimation of missing values in time series. Master’s thesis, 2006.
Ching-Lai Hor, Simon J Watson, and Shanti Majithia. Analyzing the impact of weather variables on monthly electricity demand. IEEE transactions on power systems, 20(4):2078- 2085, 2005.
Julian Faraway and Chris Chatfield. Time series forecasting with neural networks: a comparative study using the airline data. Journal of the Royal Statistical Society: Series C (Applied Statistics), 47(2):231-250, 1998.
TWS Chow and CT Leung. Neural network based short-term load forecasting using weather compensation. IEEE Transactions on Power Systems, 11(4):1736-1742, 1996.