Short Term Load Forecasting using

Deep Learning Algorithms

Project Supervisor Presented by

Prof. Vinod Kumar D M Abhishek Jaiswal(192202)

(Retired) Ishan Jain (192232)
Ashutosh Meena (192210)
Content
● Introduction
● Literature review
● Problem Statement
● Convolution Neural Network(CNN)
● Long Short Term Memory (LSTM) Network
● Simulation Steps
● Results
● Conclusion
● Future Work
● References
INTRODUCTION

Load Forecasting
● Load forecasting is a technique used by power or energy-providing companies to
predict the power/energy needed to meet the demand and supply equilibrium

● Classification of load forecasting depending on time period

○ Long Term
○ Medium Term
○ Short Term
Short Term Load Forecasting (STLF)
● It represents electric load forecast for a time interval of few minutes to a few
hours.
● STLF is important to supplier because they can use the forecasted load to control
the number of generators in operation.
○ Shut up some unit when forecasted load is low.
○ Start up of new unit load forecasted is high.
Need of Short Term Load Forecasting
● To drive the scheduling functions that decides the most economic
commitment of generation sources.
● To access the power system security based on the information available to the
dispatchers to prepare the necessary corrective actions.
● To provide the system dispatcher with the latest load predictions so that the
system can be operated both economically and reliably
● Help to estimate load flows and to make decisions that can prevent
overloading
● Grid glitch happens when high mismatch in demand and supply, potentially
due to unexpected or sudden change in power use patterns.
LITERATURE REVIEW
S. Hosein and P. Hosein, "Load forecasting using deep neural networks," 2017 IEEE Power & Energy Society
Innovative Smart Grid Technologies Conference (ISGT), 2017, pp. 1-5, doi: 10.1109/ISGT.2017.8085971.

This paper presents a new approach for short term electrical load forecasting (STLF) using deep neural networks (DNN).
They compared deep architectures and traditional methods applied to the STLF problem and provide a comprehensive
analysis of numerous deep learning models.Recurrent Neural Network-Long Short term Memory (RNN-LSTM)
technique was used whose training and running period was slower. The Mean Absolute Percentage Error (MAPE) was
2.36.

Ming-Guang Zhang, "Short-term load forecasting based on support vector machines regression," 2005
International Conference on Machine Learning and Cybernetics, 2005, pp. 4310-4314 Vol. 7, doi:
10.1109/ICMLC.2005.1527695.
In this paper, a novel method based on Support Vector Machines (SVM) regression for the electric power system
short-term load forecasting is presented. SVM is a novel powerful machine learning method based on Statistical
Learning Theory (SLT). SVM algorithm improves both training time and accuracy of the result but this method is not
suitable for large number of datasets. The MAPE was 5.6% using this method.
N. Kamel and Z. Baharudin, "Short term load forecast using Burg autoregressive technique," 2007
International Conference on Intelligent and Advanced Systems, 2007, pp. 912-916, doi:
10.1109/ICIAS.2007.4658519.
This paper proposed a technique for STLF by using the approach of signal modeling to predict one week ahead of
the future load. The proposed technique introduced a signal modeling by using Burg’s method in solving load
forecast for load demand. The method is one of the digital signal processing approach, where a signal (historical
data) PS(n) is known over a given interval of time and the goal is to predict PL(n) over some other interval.
.Autoregressive method was used to predict and the MAPE was 3.9. The past of the the time-series is not
enough to predict the future. Multiple additional features should be taken into account to get good forecasts .

S. H. Rafi, Nahid-Al-Masood, S. R. Deeba and E. Hossain, "A Short-Term Load Forecasting Method Using
Integrated CNN and LSTM Network," in IEEE Access, vol. 9, pp. 32436-32448, 2021, doi:
10.1109/ACCESS.2021.3060654.

In this paper a new approach is proposed is based on the integration of convolutional neural network (CNN) and
long short-term memory (LSTM) network. The method is applied to Bangladesh power system to provide
short-term forecasting of electrical load.. It is found that the proposed strategy results in higher precision and
accuracy in short-term load forecasting.The proposed CNN-LSTM model can deal with the long sequence time
series electric load data and predict the future load demand over a considerable period. The MAPE was 2.31
C. Liu, Z. Jin, J. Gu and C. Qiu, "Short-term load forecasting using a long short-term memory network,"
2017 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT-Europe), Turin, Italy,
2017, pp. 1-6, doi: 10.1109/ISGTEurope.2017.8260110.
In this paper, the architecture and training procedure of a LSTM network useful for short-term load forecasting is
presented. The proposed model is capable of forecasting next 24-hour load profile.The load time series studied in
this paper were recorded over a whole year, every week from Monday to Sunday at 60 minute intervals without
dealing with the load with regular and irregular jitters.

M. S. Hossain and H. Mahmood, "Short-Term Load Forecasting Using an LSTM Neural Network," 2020
IEEE Power and Energy Conference at Illinois (PECI), Champaign, IL, USA, 2020, pp. 1-6, doi:
10.1109/PECI48348.2020.9064654.
In this paper, two forecasting models using long short term memory neural network (LSTM NN) are developed to
predict short-term electrical load. The first model predicts a single step ahead load, while the other predicts
multi-step intraday rolling horizons. The time series of the load is utilized in addition to weather data of the
considered geographic area. Moreover, the performance of the LSTM NN, was compared with the performance
of other machines, namely a generalized regression neural network (GRNN) and an extreme learning machine
(ELM).
PROBLEM STATEMENT

Aiming at the problem of low accuracy of forecasting using traditional methods,

a short-term load forecasting model based on Deep Learning Networks using
Convolutional Neural Network (CNN) and a Long Short Term Memory (LSTM)
networks is proposed to predict the peak load forecast for the next hour and to
evaluate the goodness of fit of these algorithms using Mean Absolute
Percentage Error (MAPE).
Convolution Neural Network
Convolution Layer
Dense Layer and Dropout
● A neural network contains layers of
interconnected nodes. Each node is a known as
perceptron and is similar to a multiple linear
regression. The perceptron feeds the signal
produced by a multiple linear regression into an
activation function that may be nonlinear.

● Dropout works by randomly setting the outgoing

edges of hidden units (neurons that make up
hidden layers) to 0 at each update of the training
phase.

● Dropout works by randomly setting the outgoing

edges of hidden units (neurons that make up
hidden layers) to 0 at each update of the training
phase.
Long Short Term Memory (LSTM) Neural Network

Long Short Term Memory (LSTM) neural networks is a special type of Recurrent Neural Network, a
time recursive sequence neural network, which can solve the long-term dependence problem of
general Recurrent Neural Network, that is, LSTM is suitable for processing and prediction events with
very long intervals and delays in a time series.
FORGET GATE

INPUT GATE

OUTPUT GATE

LSTM Cell
 Activation function Used in LSTM

Sigmoid Activation Function Tanh Activation Function

f(x)
f(x)

input(x)
input(x)
Simulation Steps
The steps of the proposed methodology are briefly described as follows:
Step-1 Data Framing
Step-2 Constructing multi step time series.
Step-3 Building forecasting Model
Sep-4 Training Proposed Model and Validating it
Step-1: Data Framing
1. Data Used: The data is the most
important key for Machine learning.
2. Data we used is a Panama’s Hourly
load.
3. Data framing include all the data
preprocessing part.
a. Removing null values
b. Normalizing the data
4. The missing values need to be
adjusted with the trend of the load
pattern obtained from previous
record.
5. Then, a load dataset has to be divided
into training and test sets for
evaluating the proposed model.
Step-2: Constructing Multi Step Time Series
Case I: Construction of Multi Step Time Series for Convolution Neural Network
● In the proposed model, electric load dataset has to be transformed in the shape of [time steps,
features,Week].
● For training we are having 36540 samples and for testing we are having 11004 samples.
● In each sample we have three channels similar to 3 RGB channel of images and each channel the
input have a data of the 168 hours of the week
● The output expected from the model is to predict the load of next few hours.
Case II: Construction of Multi Step Time Series for Long-Short term Memory Network
● In the proposed model, electric load dataset has to be transformed in the shape of [Time steps,
features].
● For training we are having 36540 samples and for testing we are having 11004 samples.
● In each cell of LSTM Network we give 19 features and for next time stamp prediction we are using
504 cells
● The output expected from the model is to predict the load of next few hours.
Step-3: Building Forecasting Model
Case I-Convolution Neural Network Model Construction:
Input Size=(168, 19,3) Model Summary:
● There are 11 Convolution layer:
○ Kernel Size=(5,2,3) Feature Maps=15 Convolution Stride=(1,1)
○ Kernel Size=(5,2,15) Feature Maps=30 Convolution Stride=(1,1)
○ Kernel Size=(5,2,30) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,2,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,2,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,2,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,2,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,2,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,2,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,3,60) Feature Maps=60 Convolution Stride=(1,1)
○ Kernel Size=(5,3,60) Feature Maps=60 Convolution Stride=(1,1)
● One layer to flatten the 3d data to 1d data.
● Neural network with 1 hidden layer with 128 Perceptron and Dropout =0.3
Step-3: Building Forecasting Model
Case II-Long-Short Term Memory network Model
Construction:
Input Size=(504, 19) Model Summary:
● There are 4 LSTM Cells:
○ Features Unit=64, Return Sequence=True
○ Features Unit=64, Return Sequence=True
○ Features Unit=64, Return Sequence=True
○ Features Unit=64, Return Sequence=False
● Each LSTM Cell has a input of features from a hour.
● In last a Back Propagation Network to get a output.
Step-4: Training Proposed Model and Validating it

● Loss Function Used for Training purpose is Mean Absolute Error.

● Adaptive Moment Estimation is an algorithm for optimization technique for gradient descent.
Case I-Convolution Neural Network training
Step-4: Training Proposed Model and Validating it

Case II-Long-Short Term Network Training

 Results: Training Dataset Testing Dataset

CNN for Next 1 hour

Prediction

% Error=1.54% %Error=1.76%

LSTM Network for

next 1 hour
Prediction

%Error=2.59% %Error=3.07%
Results: Training Dataset Testing Dataset

CNN for Next 2 hour

Prediction
(1st hour data
result)
% Error=3.15% %Error=3.33%

CNN for Next 2 hour

Prediction
(2nd hour data
result)
%Error=3.01% %Error=3.34%
Conclusion
● Short term load forecasting is an integral part of power system planning and
operations.
● Precise forecasting of load is essential for unit commitment, capacity planning,
network augmentation and demand side management.
● High accuracy model can be used to control the power system architecture in
advance.
● Designed a CNN model and LSTM model and compared there results.
● Designed a CNN model form multiple time stamp prediction.
 FUTURE WORK
● Optimising the error percentage.
● Comparing trained models for different data sets.
● To use this model for prediction of multiple step ahead.
● Trying to develop a model with a properties of both CNN
and LSTM.
References
Source of historical load data

https://www.kaggle.com/datasets/ernestojaguilar/shortterm-electricity-load-forecasting-panama

S. H. Rafi, Nahid-Al-Masood, S. R. Deeba and E. Hossain, "A Short-Term Load Forecasting Method Using Integrated CNN and LSTM
Network," in IEEE Access, vol. 9, pp. 32436-32448, 2021, doi: 10.1109/ACCESS.2021.3060654.

M. Voß, C. Bender-Saebelkampf and S. Albayrak, "Residential Short-Term Load Forecasting Using Convolutional Neural Networks," 2018
IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm), 2018, pp.
1-6, doi: 10.1109/SmartGridComm.2018.8587494.

M. Alhussein, K. Aurangzeb and S. I. Haider, "Hybrid CNN-LSTM Model for Short-Term Individual Household Load Forecasting," in IEEE
Access, vol. 8, pp. 180544-180557, 2020, doi: 10.1109/ACCESS.2020.3028281.

A. A. Mamun, M. Sohel, N. Mohammad, M. S. Haque Sunny, D. R. Dipta and E. Hossain, "A Comprehensive Review of the Load Forecasting
Techniques Using Single and Hybrid Predictive Models," in IEEE Access, vol. 8, pp. 134911-134939, 2020, doi:
10.1109/ACCESS.2020.3010702.

N. Singh, C. Vyjayanthi and C. Modi, "Multi-step Short-term Electric Load Forecasting using 2D Convolutional Neural Networks," 2020
IEEE-HYDCON, 2020, pp. 1-5, doi: 10.1109/HYDCON48903.2020.9242917.
M. S. Hossain and H. Mahmood, "Short-Term Load Forecasting Using an LSTM Neural Network," 2020 IEEE Power and Energy
Conference at Illinois (PECI), Champaign, IL, USA, 2020, pp. 1-6, doi: 10.1109/PECI48348.2020.9064654.
C. Liu, Z. Jin, J. Gu and C. Qiu, "Short-term load forecasting using a long short-term memory network," 2017 IEEE PES Innovative Smart
Grid Technologies Conference Europe (ISGT-Europe), Turin, Italy, 2017, pp. 1-6, doi: 10.1109/ISGTEurope.2017.8260110.