PM-Gati Shakti A Case Study of Demand Forecasting
PM-Gati Shakti A Case Study of Demand Forecasting
PM-Gati Shakti A Case Study of Demand Forecasting
Demand Forecasting
*
SujayKumar Reddy M and Gopakumar G
doi: 10.20944/preprints202308.0535.v1
Power; Delhi
Copyright: This is an open access article distributed under the Creative Commons
Attribution License which permits unrestricted use, distribution, and reproduction in any
Disclaimer/Publisher’s Note: The statements, opinions, and data contained in all publications are solely those of the individual author(s) and
contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting
from any ideas, methods, instructions, or products referred to in the content.
Case Report
PM-Gati Shakti: A Case Study of Demand Forecasting
SujayKumar Reddy M 1, * and GopaKumar G 2
1 Vellore Institute of Technology, Vellore
2 Lead Contact
* Correspondence: [email protected]
Introduction
The Indian Energy GRID is maintained by POWERGRID [18] which has the objectives of running
the GRID efficiently and installing transmission lines etc... and the second one is the National Load
Dispatch Center (NLDC) [14] which concentrates on Supervision over the Regional Load Dispatch
Centres, Scheduling and dispatch of electricity over inter-regional links in accordance with grid
standards specified by the Authority and grid code specified by Central Commission in coordination
with Regional Load Dispatch Centres, Monitoring of operations and grid security of the National Grid,
etc... This research mainly focusses on this NLDC which is a Division of Ministry of Power.
Figure 1 [15] describes the Yearly Installed Power Capacity in Delhi. The highest installed Capacity
was 8,346.72 MW in Fiscal year 2016. which is responsible for sending the energy from Stations to
sub-stations and to discuss and then to homes, industries, commercials, etc... As of 21-06-2023 The
installed Capacity Sector-wise data [16] gives an overview of what type of Thermal Plants which are
present in Delhi and also what types of Energy resources are present in Delhi. There are 11 Thermal
Stations in Delhi with 4 400KV Substations and 42 200 KV Substations [17].
2 of 12
3 of 12
The correlation between these variables with the Consumption of Electricity of Delhi [19] Now, as
we speak Consumption by the consumers is based on the Maximum Demand which has been recorded
per day in the National Load Dispatch Center (NLDC) As all these entities are highly correlated
we push our limits to only use univariant analysis of Maximum Demand attained in the Daily data
produced by the Delhi Consumers.
4 of 12
Machine Learning and Time Series Forecasting models that have been previously applied to create our
Baseline model.
The "Integration of Ministries" initiative combines seven drivers: railways, ports, waterways,
logistic infrastructure, mass transport, airports, and roads. The primary goal is to enhance connectivity
and boost the competitiveness of Indian businesses. This integration is known as "Gati-Shakti," and it
rests on six pillars of Connectivity for Productivity.
1. Comprehensiveness
2. Prioritization
3. Optimization
4. Synchronization
5. Analytical
6. Dynamic
Given these Pillars, we try to formalize and study the use cases which are provided for this
problem Statement, "Maximum Demand Forecasting in Delhi", Here’s a case study to understand ...
Let us say that the Ministry of Steel[20] created an initiative to increase domestic production
that would lower dependence on imported Steel and would result in considerable savings of foreign
exchange. As this seems reasonable, there are some pros and cons related to this one being less Foreign
Exchange obviously and a con being Increased Electricity Consumption. Data explains from the
Ministry of Coal to the Ministry of Power to increase the installed Capacity or Power Generation can
be captured through Comprehensiveness.
As there’s an increase in demand for Electricity, the Ministry of Power tries to solve this problem
but as the transparency of Ministries in PM-Gati-Shakti provides to consolidate the increase in demand
like the Ministry of Railways tries to optimize routes on weekends and majorly metro trains which
are in Delhi can be optimized to consolidate this increase in demand which can be captured through
Prioritization, Optimization while maintaining a Holistic Approach. Note that this is a scenario that
can be predicted but there will be many scenarios as possible in-order to facilitate the Maximum
Demand Forecasting.
Related Work
This Section consists of Background analysis a.k.a Literature review of the Demand Forecasting
Techniques adopted and Analyzing the recent reports by the Ministries of Government. This section
focuses on many different attributes which are needed to be considered by the previous research done
by the individuals.
Deep Learning
Figure 3 (a) gives us a clear understanding of a number of Deep Learning papers which are
mostly based on Indian Authors published in IEEE. Anil et al [1] use the Levenberg-Marquardt back
propagation algorithm ANN on the day ahead Short term load Forecasting on the state of Uttar
Pradesh trained on hourly data with the MAPE score of average MAPE 3.05, This work suggests using
the ANN model to check with our dataset too. Navneet et al [2] use the New Delhi ADEL data to
forecast the load by using different Neural Network Architectures in which ELMANN Neural Network
Architecture has given good accuracy. Dharmoju et al [3] provided a sector of Residental buildings by
the United States Dataset by using LSTM (Long Short Term Memory) model for monthly forecasting.
Shaswat et al [4] use a Temporal Fusion Architecture to capture the interactions which are scaled
between 0 and 1 for daily data which achieves 4.15% more than the existing models and this is for the
whole of India which is not region specific. Saravanan et al [5] use economic factors like GDP, national
income, consumer price index, etc.. with that they used Principle component Analysis followed by
ANN which gives the highest accuracy of MAPE score of 0.43. Vishnu et al [6] concentrate on the work
on Renewable Energy Resources and devised two major LSTM (Long Short Term Memory) models.
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 August 2023 doi:10.20944/preprints202308.0535.v1
5 of 12
Machine Learning
Figure 3 (b) gives us a clear understanding of a number of Machine Learning papers and their
analysis in all the papers (as ids). Christos et al [7] which also forecasts the Peak Demand in the
electrical sector of the Producers side. They used a Netherland dataset from 3 regions and analyzes
comparatively with all the existing Forecasting methods such as ARIMA, Ridge, and Lasso Regression
and the results suggested that the Bi-directional LSTM. Saravanan et al [8] formalize a set of 64 if-else
statements and the variables include per-capita GDP, population, and Import/Export and they have
achieved the MAPE of 2.3. Mannish et al [9] devised an Ensemble Approach for the DISCOMs in
the region Delhi for the Post Covid Scenario and their proposed model includes combining XGBoost,
LightGBM, and CatBoost algorithms which achieved an average MAPE of 5.0. Banga et al [10]
compared many Machine Learning Algorithms for the dataset which considers 29 attributes and the
Facebook Prophet model outscores daily and hourly datasets with MAPE scores of 0.4 and 0.2.
6 of 12
Data Overview
The Features of the data which has been provided below
1. Date (DD/MM/YYYY)
2. Max.Demand met during the day (MW)
3. Shortage during maximum Demand (MW)
4. Energy Met (MU)
5. Drawal Schedule (MU)
6. OD(+)/UD(-) (MU)
7. Max OD (MW)
8. Energy Shortage (MU)
Energy Shortage (MU) feature is not available every day. This feature is recorded from 2017-05-09
as per the reports generated by the PSP by POSOCO. The data is available here [21]. This paper focuses
on Univariant Analysis not to make it as complex, but to consider Max Demand met during the day
(MW) as a single column.
As the column "Max Demand met During the day" is the major feature that we consider. The data
starts from 2013-04-01 to 2023-05-31 which consists of 3713 days but the data points only consist of
3640 with the missing data we use the Imputation Techniques dataset and Non-Imputation Techniques
dataset (no_null).
For Imputation Techniques, this paper considers Mean, Median, Mode, and Linear Interpolation
Imputation data that has been imputed. So, as combined this generates 5 datasets where the models
are applied to compare the performance of which Imputation is good.
Table 1 depicts the datasets which are created from the univariant data taken from one of the
features in the dataset. ("Max. Demand met during the day (MW) "). As explained in the Methodology
section the dataset is divided into train and test to take the MAPE score. For ARIMA models we
generate train MAPE and test MAPE to check the overfitting criteria also with AIC, BIC, log(p), etc...
Forecasting Models
This paper develops a Time series forecasting model ARIMA which is known to be an
Auto-Regressive Integrated Moving Average. the models which include AR, MA and ARMA, and
ARIMA, and develop a model list from these regression types using the parameters. The major
parameters included in the arima model are p, q, and d where p is the parameter for Auto-regressive
co-efficient which says about how many days have the co-relation between today’s date.
The real-world data tends to be always non-Stationary. A signal is said to be stationary if its
statistical properties like mean, standard deviation, trend, etc... doesn’t change over time. To check if
the time series is stationary or not, we use Augmented-Dickey Fuller Test where the null hypothesis is
"the time series contains a unit root and is non-stationary". The results of the Augmented Dickey-Fuller
Test for each of the Imputation datasets are given in the below sections.
7 of 12
The ADF test for Mean imputation dataset test-statistic = -5.393 p-value = 3.49e-06 for first
difference to try to change the time-series to stationary. test-statistic = -10.073 p-value = 1.23e-17 for
Second Difference, which is not suggested as the p-value is zero (over-differencing) test-statistic =
-21.617 p-value = 0.0
The ADF test for Median imputation dataset test-statistic = -5.363 p-value = 4.042e-06, for first
difference to try to change the time-series to stationary. test-statistic = -10.075 p-value = 1.223e-17 for
Second Difference, which is not suggested as the p-value is zero (over-differencing) test-statistic =
-21.686 p-value = 0.0
The ADF test for Mode imputation dataset test-statistic = -5.227 p-value = 7.73e-06 for first
difference to try to change the time-series to stationary. test-statistic = -10.258 p-value = 4.31e-18 for
Second Difference, which is not suggested as the p-value is zero (over-differencing) test-statistic =
-22.121 p-value = 0.0
The ADF test for Linear Interpolation imputation dataset test-statistic = -5.390 p-value = 3.53e-06
for first difference to try to change the time-series to stationary test-statistic = -10.072 p-value = 1.24e-17
for Second Difference, which is not suggested as the p-value is zero (over-differencing) test-statistic =
-21.44 p-value = 0.0
In this formula:
To estimate the parameters (φ1 , φ2 , . . . , φ p , θ1 , θ2 , . . . , θq ) and the intercept (c) of the ARMA(p, q)
model, various estimation techniques can be used, such as maximum likelihood estimation.
Once the parameters are estimated, the ARMA(p, q) model can be used for forecasting by
substituting the lagged values and lagged error terms of the time series into the formula to predict
future values.
Note that the ARMA(p, q) model assumes stationarity of the time series, and it is a flexible model
that can capture both autoregressive and moving average components in the data.
8 of 12
9 of 12
10 of 12
11 of 12
Table 11. SARIMA Model Comparison Results for Linear Interpolation Imputation.
In conclusion, our study highlights the importance of integrating more data into the PM
Gati-Shakti Scheme in order to validate the findings presented in the Introduction section. The base
models provide a preliminary understanding of the scheme’s potential, but further data incorporation
is crucial for robust conclusions. By expanding the scope of our analysis to encompass a wider range
of variables and factors, we can enhance the accuracy and reliability of the case study.
Furthermore, our future work will focus on employing Reinforcement Learning techniques
for model selection. This approach is particularly relevant when dealing with a larger dataset that
integrates ministries within the Union Territory of Delhi. Reinforcement Learning algorithms can
effectively evaluate and select the most suitable models by considering the complex interactions and
dependencies between different variables. By leveraging the power of machine learning and advanced
analytics, we can make informed decisions that lead to better outcomes and enhanced efficiency within
the PM Gati-Shakti Scheme.
In summary, our research emphasizes the need for data integration and the application of
Reinforcement Learning in the context of the PM Gati-Shakti Scheme. These steps will contribute
to a more comprehensive understanding of the scheme’s impact and enable evidence-based
decision-making for the integration of ministries in the Union Territory of Delhi. By continuously
improving our analytical approaches, we can enhance the effectiveness of the scheme and drive
positive socio-economic outcomes.
Acknowledgments: This project was supported by National Institute of Technology, Calicut a center of
PM-Gati-Shakti Scheme initiated by National Institute of Industrial Engineering.
References
1. Anil K Pandey, Kishan Bhushan Sahay, M. M Tripathi, and D Chandra, Short-term load forecasting of
UPPCL using ANN, 2014 6th IEEE Power India International Conference (PIICON), 2014, pp. 1-6, doi:
10.1109/POWERI.2014.7117741
2. Navneet Kumar Singh, Asheesh Kumar Singh, and Manoj Tripathy, A comparative study of BPNN, RBFNN and
ELMAN neural network for short-term electric load forecasting: A case study of Delhi region, 2014 9th International
Conference on Industrial and Information Systems (ICIIS), 2014, pp. 1-6, doi: 10.1109/ICIINFS.2014.7036502
3. Pavan Kumar Dharmoju, Karthik Yeluripati, Jahnavi Guduri, and Kowstubha Palle, Forecasting Electrical
Demand for the Residential Sector at the National Level Using Deep Learning, 2021 International Conference on
Artificial Intelligence and Machine Vision (AIMV), 2021, pp. 1-6, doi: 10.1109/AIMV53313.2021.9670956
4. Shashwat Jha and Vishvaditya Luhach, Indian Peak Power demand Forecasting: Transformer Based Implementation
of Temporal Architecture, 2022 IEEE Global Conference on Computing, Power and Communication Technologies
(GlobConPT), 2022, pp. 1-5, doi: 10.1109/GlobConPT57482.2022.9938155
5. S Saravanan, Kannan Subramanian, and C. Thangaraj, India’s Electricity Demand forecast using Regression
Analysis and Artificial Neural Networks based on principal components, ICTACT Journal on Soft Computing, vol. 2,
pp. 365-370, Jul. 2012, doi: 10.21917/ijsc.2012.0057
6. Vishnu Vardhan Sai Lanka, Millend Roy, Shikhar Suman, and Shivam Prajapati, Renewable Energy and Demand
Forecasting in an Integrated Smart Grid, 2021 Innovations in Energy Management and Renewable Resources(52042),
2021, pp. 1-6, doi: 10.1109/IEMRE52042.2021.9386524
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 August 2023 doi:10.20944/preprints202308.0535.v1
12 of 12
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those
of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s)
disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or
products referred to in the content.