2019 First International Symposium on Instrumentation, Control, Artificial Intelligence, and Robotics (ICA-SYMP)

Short-Term Load Forecast for Energy Management

Systems Using Time Series Analysis and Neural
Network Method with Average True Range
Tanwalai Panapongpakorn1 and David Banjerdpongchai2
Department of Electrical Engineering, Faculty of Engineering, Chulalongkorn University,
254 Phayathai Road, Pathumwan, Bangkok, 10330 Thailand
Email: tanwalaimind@gmail.com and 2bdavid@chula.ac.th (corresponding author)
1

Abstract— The load forecasting is one of the important tools predict future 30 minutes ahead. We employ historical load,
for Energy Management System (EMS). It is used for weather data including temperature and irradiance as the input
planning and management power balance. Short term load features for ANN and LSTM based RNN to predict load 30
forecasting (STLF) has a significant impact on the efficiency minutes ahead. Additional, we employ ATR as another input of
of operation. The load forecasting model must be able to the LSTM based RNN. Particularly, the inputs of RNN model are
accurately predict the demand of electrical power. This paper historical load, weather data and ATR index. The data is taken
proposes the load forecasting models based on time series from MHS recorded in 2013. We divide data set into three periods
analysis and neural network methods. The data is taken from and develop STLF models using different methods.
Mae Hong Son (MHS) located in the northern Thailand. Time The paper is organized as follows. In the next section, we
series analysis utilizes autoregressive integrated moving propose 5 forecast models, ARIMA, SARIMA, ANN, RNN and
average (ARIMA) and seasonal autoregressive integrated RNN with ATR models. In section 3, we present forecast
moving average (SARIMA). In addition, neural networks procedure, followed by the performance measures in section 4.
cover artificial neural network (ANN) and long-short term Section 5 we give numerical results and compare the accuracy.
memory (LSTM) based recurrent neural network (RNN). Section 6 gives discussion and section 7 concludes the results.
Additional, the Average True Range (ATR) index is adapted
to improve the performance of RNN model. We compare the II. FORECAST MODELS
performance of these models using statistic criteria, namely, A. ARIMA model
root mean square error and mean absolute percent error and The ARIMA model is time series analysis proposed by Box
choose the best model to implement for micro EMS. and Jenkins in 1970s [2]. The general form of ARIMA model
Keywords—short term load forecast, energy management systems, ARIMA (p, d, q) where p and d are orders of auto-regressive (AR)
time series analysis, artificial neural network, recurrent neural and moving average (MA), d is the order of the differences. Since
network, average true range. the ARIMA model cannot track the seasonal component in load
characteristic, the modified ARIMA called SARIMA is used for
I. INTRODUCTION
tracking the seasonal component in load time series. The general
In a smart grid, the ability of electricity utility to supply
form of ARIMA (p, d, q) shown as follows.
sufficiently power meet the demand in area is important. Short
୮ ୯
term load forecast (STLF) is one of important tools of energy ൫ͳ െ ෌୩ୀଵ Ƚ୩ ୩ ൯ሺͳ െ ሻୢ ›୲ ൌ ൫ͳ ൅ ෌୩ୀଵ ୩ ୱ୩ ൯ɂ୲ (1)
management systems (EMS) to support power and energy
balance in the area. The load forecast is used for planning and where yt is actual load at time t, μ is constant, Ƚ୩ is AR
controlling the startup and shut down of power generation units. coefficient, ୩ is MA coefficients, ɂ୲ is error term at time t.
There are several methods for STLF such as regression methods, B. SARIMA model
time series models, fuzzy logic, neural networks, expert system, The SARIMA model has general form as follows
and hybrid approaches [1]. The Average True Range (ATR) is SARIMA (p, d, q) (P, D, Q) s, P, D, and Q are the seasonal
one of the statistical methods which calculates the volatility in order and S is the step of seasonal differences.
stock market price. This can be used for forecast model. ୮ ୔
In this paper, we focus on STLF which estimates demand in
ቌͳ െ ෍ Ƚ୩ ୩ ቍ ൭ͳ െ ෍ Ԅୱ୩
୩ై
൱ ሺͳ െ ୱ ሻୈ ሺͳ െ ሻୢ ›୲
future of 30 minutes. In particular, we compare four different
୩ୀଵ ୩ୀ
models for STLF. There are time series models, namely,
୕ ୯
Autoregressive Integrated Moving Average (ARIMA) and ൌ ቀͳ ൅ ෌୩ୀଵ Ԅୱ୩ ୱ୩
୐ ቁ ൫ͳ ൅ ෌୩ୀଵ  ൯ɂ୲ (2)
Seasonal Autoregressive Integrated Moving Average (SARIMA)
models. In addition, there are artificial intelligent models, namely, C. Feed Forward ANN
Feed Forward Neural Network (ANN), Long-Short Term The architecture of ANN model is multilayer perceptron
Memory based Recurrent Neural Network (LSTM based RNN) (MLP) that consists of three layers, an input layer, a hidden
and Long-Short Term Memory based Recurrent Neural Network layer and an output layer. The number of neurons in input
(LSTM based RNN) with Average True Range (ATR) models. layer depends on input features following the studied cases
For time series models, the only historical electricity consumption and number of neurons in hidden layer.
data is used for developing ARIMA and SARIMA models which

978-1-5386-7774-2/19/$31.00 ©2019 IEEE

86

Authorized licensed use limited to: COVENTRY UNIVERSITY. Downloaded on January 02,2024 at 05:03:21 UTC from IEEE Xplore. Restrictions apply.
 2019 First International Symposium on Instrumentation, Control, Artificial Intelligence, and Robotics (ICA-SYMP)
D. RNN It makes model change to SARIMA model. After stationary
The LSTM is one type of RNN which can solve the process, the time series data can be used for building models.
gradient vanishing problem that occurs for prediction the long
Step 2: Model identification
sequence time series. Single unit LSM makes decision by
We determine orders of ARIMA ሺ’ǡ †ǡ “ሻ and SARIMA
considering current input and previous output and memory. It
ሺ’ǡ †ǡ “ሻሺǡ ǡ ሻୱ  models using autocorrelation function (ACF)
generates a new output and alters its memory. The xt is input at
and partial autocorrelation function (PACF) of data at lag k.
time t, yt is output at time t, and ht is hidden state at time t.
We perform a periodic data every 48 samples and determine the
E. RNN with ATR
order of AR, MA, SAR and SMA.
The RNN with ATR model is the same as the LSTM based
To achieve the best models for ARIMA and SARIMA, the
RNN model. However, the special characteristic is that ATR
model with lowest Akaike information criterion (AIC) and
index is the additional input of the RNN model. ATR index of
Bayesian information criterion (BIC) is chosen. The ARIMA
the forecast day is calculated by
and SARIMA models which have lowest AIC and BIC in each
ATR = (ൣσ୒ ୧ୀଵሺୢି୧ െ ୢି୧ ሻ൧Ȁ period total tens models (five for ARIMA and five for SARIMA)
where ୢି୧ is peak load, ୢି୧ is valley load, d is forecasted are estimated parameters in next step.
day, and N is the number of historical data. In this study, we
use ATR which is calculated by using load data from 14 Step 3: Parameter Estimation
previous days. Estimate parameters of chosen ARIMA and SARIMA
models using the maximum likelihood estimation method.
III. FORECAST PROCEDURE
Step 4: Diagnosis the fitness of model
A. Data collection Using Ljung-Box Q statistic to check the adequacy. The
Data is the historical load and weather data from Mae Hong residuals must be normally distributed, if not-normally
Son year 2013. The data was divided into three periods: The first distributed go to step 2. Otherwise, proceed to step 5.
is 1 May to 30 June, the second period is 1 July to 30 September Step 5: Forecasting and Validation
and the last period is 1 October to 31 December. After hypothesis test of parameters of all models (ten models).
B. ARIMA and SARIMA models The ARIMA 5 models and SARIMA 5 models in each period
Each data set were taken to develop ARIMA and SARIMA are evaluated by validation data set.
model were processed for building time series model base Box- The validation data set of each period is used to forecast the
Jenkins method. The flowchart for developing the models and 30 minutes ahead load. The validation data set is transformed to
evaluating the model’s performance as shown in Fig. 1. The natural log scales same as the training data set.
process can be explained in 4 steps Each model used to forecast one-step ahead load evaluated by
the validation set, then the error of forecast from each model are
used to assess the forecasting accuracy criteria. The criteria chosen
to measure accuracy of forecast in this paper is RMSE and MAE.
C. ANN, RNN and RNN-ATR models
The ANN, RNN and RNN with ATR models have been
implemented to forecast load at time t by knowing the historical
series of electricity consumption and weather data. The collection
of different type of data: 1. Weather data: temperature (T) and
irradiance (I) and 2. Historical series of peak values of load (L) in
time step 30 minutes. The data set which is weather and load data
in MHS is divided into training data set and validation data set for
each period that we mentioned in data collection. All data set are
normalized to be in range -1 to 1 before training and testing the
model. There are three studied cases of ANN and LSTM based
RNN divided by number of input variables. The input variables for
ANN models provide by stepwise regression method to find the
significant variables to dependent variable at time t. From stepwise
regression method we can provide 3 cases as shown in Table I.

TABLE I. STUDIES OF ANN AND RNN MODELS

Input Variables
Model
Temperature Irradiance Load
1 Tt-2 It-2 Lt-2, Lt-1
Fig. 1 Flow chart for development and evaluating performance of
time series models. 2 Tt-1 It-1 Lt-3, Lt-2, Lt-1
Step 1: Preprocessing data 3 Tt-3, Tt-1 It-3, It-1 Lt-3 Lt-2, Lt-1
Transform all data into natural log scale and diagnostic for
stationary by using Augmented Dickey Fuller (ADF) test. The Each model consists of 10 different topologies of network
testing result show that original data is non-stationary. Thus, we (number of neurons in hidden layer). The topology defines ANN
apply the differencing method of non-seasonal integrated term. case 1 (number of neurons in input layer, number of neurons in
However, it is not enough to explain the seasonality in time series hidden layer). The ANN and RNN models are trained by using
data. We apply the seasonal integrated model to time series data. Levenberg-Marquardt optimization algorithm. Note that Tt is

87

Authorized licensed use limited to: COVENTRY UNIVERSITY. Downloaded on January 02,2024 at 05:03:21 UTC from IEEE Xplore. Restrictions apply.
 2019 First International Symposium on Instrumentation, Control, Artificial Intelligence, and Robotics (ICA-SYMP)
temperature at time t, It is irradiance at time t, and Lt is load 2013. For ARIMA ሺ’ǡ †ǡ “ሻ in the first period, we vary ’ from
at time t. The development procedure for all ANN and RNN 1 to 5, d from 1 to 2 and q from 1 to 4. For SARIMA
models can be shown in Fig. 2. ሺ’ǡ †ǡ “ሻሺǡ ǡ ሻୱ model, we vary ’ǡ †ǡ “ same as the ARIMA
After training ANN and RNN models, the mean square error model, from 1 to 8,  from 1 to 3 and D from 1 to 2.
(MSE) and statistic for measurement of goodness model’s fit R- There are totally 40 ARIMA models and 1920 SARIMA models.
squared are criterion for selection models. The ANN model For ARIMA ሺ’ǡ †ǡ “ሻ in the second period, we vary ’ from
which has lowest MSE and highest R-squared values from the 1 to 8, † from 1 to 2, “ from 1 to 4. For SARIMA
training process is selected to implement for one-step ahead load ሺ’ǡ †ǡ “ሻሺǡ ǡ ሻୱ model, we vary  from 1 to 5,  from 1 to 4
forecasting. The three chosen ANN models are used to forecast and D from 1 to 2. There are totally 64 ARIMA models and 2560
load. SARIMA models.
After we choose the RNN model in each period, the chosen For ARIMA ሺ’ǡ †ǡ “ሻ in the last period, we vary ’ from 1 to
ATR index is integrated to the RNN models, called “RNN-ATR 9,† from 1 to 2, “from 1 to 5. For SARIMA ሺ’ǡ †ǡ “ሻሺǡ ǡ ሻୱ
model”. For RNN-ATR model, ATR index is added together model, we vary  from 1 to 8,  from 1 to 5 and D from 1 to 2.
with other historical inputs, namely, temperature, irradiance and There are totally 90 ARIMA models and 7200 SARIMA models.
load. The load for each validation set from previous 14 days is To evaluate the performance of data fitting of each model,
used to calculate ATR index for the forecast day. ATR index is we evaluate using validation data set in each period. After
normalized as the other inputs. forecasting load 30 minutes ahead, all proposed models are
evaluated by RMSE and MAE as shown in Table II, III, and III.

TABLE II. PERFORAMNCE INDICES USING TRAINING DATA

AND VALIDATION DATA IN THE FIRST PERIOD
Using Training Data Using Validation Data
Model RMSE MAE RMSE MAE
kW % kW % kW % kW %
ARIMA
(3, 2, 1) 301.05 7.76 265.16 6.83 305.08 9.32 275.17 8.41
SARIMA
(3, 2, 1) 299.85 7.72 264.89 6.82 301.75 9.22 270.87 8.28
(2, 2, 3)48
ANN
(5, 20)
290.68 7.49 245.52 6.32 299.67 9.16 250.14 7.64
RNN
(5, 16) 285.47 7.35 227.56 5.86 291.44 8.91 229.36 7.01
RNN-ATR
(5, 16) 287.12 7.40 221.52 5.71 289.12 8.84 228.39 6.98

TABLE III. PERFORMANCE INDICES USING TRAINING DATA AND

VALIDATION DATA IN THE SECOND PERIOD
Using Training Data Using Validation Data
Model RMSE MAE RMSE MAE
kW % kW % kW % kW %
ARIMA
(7, 2, 1)
302.54 9.25 270.85 8.28 303.51 7.03 274.72 6.36
SARIMA
(7, 2, 1) 299.45 9.15 265.85 8.12 300.14 6.95 269.61 6.24
(1, 2, 1)48
ANN
(5, 10)
292.78 8.95 239.63 7.32 293.92 6.80 240.45 5.57
RNN
(5, 20)
251.67 7.69 181.64 5.55 252.61 5.85 184.75 4.28
RNN-
ATR 232.40 7.10 182.07 5.56 234.07 5.42 183.05 4.24
Fig. 2 Flow chart for development of ANN and RNN models. (5, 20)

D. Performance Indices TABLE IV. PERFORMANCE INDICES USING TRAINING DATA

AND VALIDATION DATA IN THE LAST PERIOD
If the models do not perform up to the acceptable level, it can Using Training Data Using Validation Data
be substituted by the benchmark model itself (Gilliland & Model RMSE MAE RMSE MAE
kW % kW % kW % kW %
Sglavo, 2010). In this study, we employ mean absolute error
ARIMA
(MAE) and root mean squared error (RMSE) as the performance (5, 2, 2)
301.84 8.65 267.35 7.66 302.39 7.05 269.83 6.25
SARIMA
indices for forecast accuracy. They are expressed as (5, 2, 2) 302.74 8.67 267.84 7.67 303.85 7.03 268.36 6.21
ଵ
MAE = σ୬୲ୀଵห୲ െ ෡୲ ห (3) (4, 2, 2)48
ANN
୬ 296.41 8.49 240.73 6.90 297.76 6.89 242.23 5.61
(5, 22)
ଵ
 ൌ ට σ୬୧ୀଵሺ୧ െ ෡న ሻଶ (4) RNN
(5, 18)
274.54 7.87 234.94 6.78 276.55 6.40 235.35 5.45
୬
RNN-
where ෡୲ is the forecast, ୲ is actual value and n is the number ATR 270.89 7.76 215.07 6.16 273.38 6.33 217.08 5.02
(5, 18)
of observations.
From Table II, III and IV, for first period, RNN with ATR (5,
IV. NUMERICAL RESULTS 16) outperforms other models which has lowest RMSE 289.125
A. Model Accuracy Comparison kW and MAE 228.395 kW. For second period, RNN with ATR
In first data set, the training start from 1 May to 11 June 2013 (5, 20) outperforms other models which has lowest RMSE
and validation set start from 12 June to 30 June 2013. In second 234.078 kW and MAE 183.057 kW. For last period, RNN with
data set, the training start from 1 July to 2 September 2013 and ATR (5, 16) outperforms other models, which has lowest RMSE
validation set start from 3 September to 30 September 2013. In 273.388 kW and MAE 217.080 kW. It seems that when we add
last data set, the training start form 1 October to 2 December the ATR index to be input of RNN model, the forecast results are
2013 and validation set start from 3 December to 31 December better than models which do not use ATR index.

88

Authorized licensed use limited to: COVENTRY UNIVERSITY. Downloaded on January 02,2024 at 05:03:21 UTC from IEEE Xplore. Restrictions apply.
 2019 First International Symposium on Instrumentation, Control, Artificial Intelligence, and Robotics (ICA-SYMP)
V. DISCUSSIONS
B. One-Step Ahead Load Forecasting There are five models for one-step ahead load forecast. We
The numerical results for one-step ahead load forecast by evaluate the performance of load forecast using validation
five candidate models are presented in Fig. 3, 4 and 5. For data set. For first period, RNN with ATR (5, 16) outperforms
first period, the validation data consists of 48 observations on other models with the minimum RMSE of 289.125 kW and
18 June 2013, the second period the data is taken from 15 MAE of 228.395 kW. For second period in validation set,
September 2013 and the last period the data is taken from 15 RNN with ATR (5, 20) outperforms other models to give the
December 2013. The observations start from 0.00 to 23.30. minimum RMSE of 234.078 kW and MAE of 183.057 kW.
The time step for load forecast is 30 minutes. In the last period, RNN with ATR (5, 18) outperforms other
models with the minimum RMSE 273.388 kW and MAE
10 6
30 Minutes ahead Load Forecasting by 5 Models on 15 June 2013 in MHS
4
Actual
ARIMA
217.080 kW. In summary, we may choose the forecast model in
ANN
SARIMA three seasons. RNN with ATR (5, 16) is chosen to forecast load
3.5 LSTM
LSTM with ATR
in May to June. RNN with ATR (5, 20) is selected to forecast
3
in July to September. RNN with ATR (5, 18) is chosen to
Load (W)

forecast in October to December.

2.5

VI. CONCLUSIONS
2

For this paper, we study the performance of forecasting

1.5
models for one-step ahead (30 minutes) load prediction in
0 5 10 15 20 25 30 35 40 45 50

Time MHS. The models which we study are time series models
Fig. 3. Forecasting load of one-day horizon on June 15, 2013. ARIMA, SARIMA models and machine learning models are
ANN, RNN and RNN-ATR models. In this study, we develop
30 Minutes ahead Load Forecasting by 5 Models on 15 September 2013 in MHS
4.5
10 6

Actual load forecast models in each period. The models for one-step
ANN
4
ARIMA
SARIMA
ahead load prediction in each period are time series models,
LSTM with ATR

3.5
LSTM i.e., ARIMA, SARIMA models and machine learning models,
i.e., ANN and RNN models. After we evaluate the models
Load (W)

with acceptable performance in each season of year, we will

2.5
implement the chosen models to load forecast module in
2
micro-EMS. For the ongoing work, we will further improve
1.5 load forecast module for better applications in microgrid to
1
make micro-EMS performs well.
0 5 10 15 20 25 30 35 40 45

Time

Fig. 4. Forecasting load of one-day horizon on September 15, 2013. ACKNOWLEDGEMENT

The first author thanks Electricity Generating Authority of
30 Minutes ahead Load Forecasting by 5 Models on 15 December 2013 in MHS
3.5
10 6

Actual Thailand for providing Master scholarship to graduate study

ANN
ARIMA
SARIMA
program at Dept. of Electrical Engineering, Chulalongkorn
3 LSTM with ATR
LSTM University.
2.5

REFERENCES
Load (W)

2
[1] I. S. Msiza, F. V. Nelwamondo, T. Marwala, “Artificial
neural networks and support vector machines for water
1.5

demand time series forecasting,” In Proc. IEEE

1
0 5 10 15 20 25 30 35 40 45
International Conf. on Systems, Man and Cybernetics,
Time
2007, Oct. 2007, pp. 638-643.
Fig. 5. Forecasting load of one-day horizon on December 15, 2013. [2] H. R. Maier, G. C. Dandy, “Neural networks for the
From figure. 3, 4, 5, we focus on the load forecast on the prediction and forecasting of water resources variables: a
days which do not perform well. It is observed that RNN and review of modelling issues and applications,”
RNN-ATR models perform better than other models for three Environmental Modelling & Software, vol. 15, no. 1, pp.
seasons. The ARIMA models perform worse for all seasons 101-124, 2000.
because the ARIMA models are particularly simple models
[3] J. Vermaak and E. C. Botha, “Recurrent Neural
which are essentially linear update models with noise
Networks for Short-Term Forecast of the Power System
interference. On the other hand, the ANN and RNN models
in Kuwait”, Applied Mathematical Modeling, vol. 21,
are nonlinear time series model. With sufficient data set,
pp. 215-219, 1997.
ANN and RNN models can match the nonlinearities of real
data well, which makes the load prediction much more [4] I. S. Msiza, F. V. Nelwamondo, and T. Marwala,
accurate than the ARIMA and SARIMA models. The RNN “Artificial neural networks and support vector machines
performs better than ANN for all seasons because the RNN for water demand time series forecasting”. In Systems,
model is best for all types of sequential data analysis. As in Man and Cybernetics, IEEE International Conference,
forecast data change with time, the RNN can learn changes in Oct. 2007. pp. 638-643.
time domain, it leads to better solutions for prediction. [5] W. Jacobs, A. Mendonca Souza and R. Ruviaro Zanini,
Moreover, the RNN-ATR model performs better than other "Combination of Box-Jenkins and MLP/RNA models for
models because the ATR index explains the characteristic of forecasting", IEEE Latin America Transactions, vol. 14,
volatility in load on forecasted day. The RNN-ATR model no. 4, pp. 1870-1878, 2016
performs better than the RNN without ATR.

89

Authorized licensed use limited to: COVENTRY UNIVERSITY. Downloaded on January 02,2024 at 05:03:21 UTC from IEEE Xplore. Restrictions apply.