Page 221 - 2024-Vol20-Issue2

217 |
forecasting.
Overall, the comparative analysis of Odisha and Delhi states
using the hybrid 1D CNN-BI LSTM model demonstrated its
effectiveness in accurately predicting peak load values. The
model’s architecture, training options, and the obtained results validation set contained 10% of the data, consisting of 142
pro- vide valuable insights for future research and practical
applications in the field of electricity forecasting. This sec-
tion presents a comparative analysis of the long-term electric-
ity forecastingmodels conducted for the states of Orissa and
Delhi. The data used for the analysis was collected from the
Odisha Power Transmission Corporation Limited (OPTCL)
forOrissa and the Northern Regional Load Dispatch Centre
(NRLDC) for Delhi. The analysis focused on evaluating the
differences in data collection, variables collected, data clean- regions necessitate customized load forecasting models. On
ing, data transformation, data reduction, forecasted response,
forecasting models, and the training, validation, and testing
sets. The duration and frequency of historical data collection
varied across the twostates, which can impact the accuracy of
load forecasting models. In Odisha, electricity load data were
collected every hour for a period of five years (January 2018
to December 2022). On the other hand, in Delhi, electricity
load data was collected every 30 minutes for a period of two
years (January 2019 to December 2022). It is important to
note that the difference in data collection intervals and dura-
tions can influence the accuracy of load forecasting models,
and thus, careful consideration must be given to selecting an
appropriate time interval and duration for data collection. The
variables collected for both states included Wet Bulb Temper-
ature, Precipitation, Dew/Frost Point, Normal Temperature,
Relative Humidity, Specific Humidity, and Wind Speed. Data
cleaning and transformation were performed for both datasets
to resolve any missing values and integrate multiple files into
a usable format. Additionally, data reduction techniques were
applied to capture the essential properties of the data while
removing redundancies. The forecasted response for Odisha
was the load demand (in MW) for a durationof 12 months,
whereas, for Delhi, it wasthe load demand (in MW) for a
duration of 9 months. Regression analysis and a hybrid deep
learning model utilizing a 1D CNN BI LSTM architecture
were employed for load forecasting in Orissa. In contrast,
Delhi’s load forecasting models included Artificial Neural
Networks (ANN), Long Short-Term Memory (LSTM), Ex-
ponential Gaussian Process Regression (GPR), and a hybrid
deeplearning approach. The training, validation, and testing
set for Odisha comprised a time interval of 1 hour, with a total posed to address uncertainty in electricity forecasting models,
of 1,819 rows and 6 columns. The training set consisted of
70ofthe data, corresponding to 1,273 rows from January 2018
to June 2021 (42 months). The validation set contained 10%
of the data, consisting of 181 rows from July 2021to Decem-
ber 2021 (6 months). The remaining 20% of the data formed
the testing set, with 363 rows from January 2022 to December
Gochhait, Sharma & Bachute
2022 (12 months). For Delhi, the time interval was set to
day-wise, resulting in a total of 1,425 rows and 6 columns.
The training set constituted 70% of the data, comprising 997
rows from January 2019 to August 2021 (32 months). The
rows from September 2021 to December 2021 (4 months).
The testing set included20% of the data, with 285 rows from
January 2022 to September 2022 (9 months). Overall, this
analysis reveals differences in dataset size, the time interval
for data collection, and forecasted response between Odisha
and Delhi. These variations canbe attributed to the unique en-
ergy demand patterns and requirements of each state. Odisha’s
diverse landscape and varying load patterns across different
the other hand, Delhi’s high population density and a large
number of commercial and industrial establishments deman-
daccurate load forecasting to effectively manage its electricity
distribution and ensure reliable power supply.
V. CONCLUSION
In conclusion, this comparative analysis of long-term elec-
tricity forecasting in Odisha and Delhi has provided valuable
insights into their load forecasting models, dataset characteris-
tics, and forecasted responses. The study revealed that Odisha
exhibitsdiverse load patterns across different regions, while
Delhi’s high population density and commercial/industrial
establishments impact load forecasting. The load forecast-
ing models employed in Odisha included regression analysis
anda hybrid deep learning approach, whereas Delhi utilized
Artificial Neural Networks (ANN), Long Short-Term Mem-
ory (LSTM), Exponential GPR, and a hybrid deep learning
approach. Notably, the forecasted response differed between
the two states, with Odisha focusing on load demand for 12
months and Delhi considering a 9-month duration. These
variations in dataset size, time intervals, and forecasted re-
sponse reflect the unique energy demand patterns observed in
each state. Short-term load forecasting presented challenges
for both states, particularly due to the volatile nature of com-
mercial and industrial consumers in Maharashtra and Delhi.
Telangana’s rapid growth and investments in smart grid tech-
nologies influenced load forecasting, while Odisha’s diverse
consumption patterns necessitated efficient short-load fore-
casting models. Additionally, a statistical framework was pro-
benefiting Load Dispatch Centers in Maharashtra, Telangana,
Odisha, and Delhi. This framework considers factors such
as high demand, rapid growth, diverse consumption patterns,
and high population density. In summary, this research paper
introduced a novel hybrid deep learning model forlong-term
load forecasting in power systems management. The model,
incorporating afeature extraction module, a densely connected