IRJET- Wind Energy Storage Prediction using Machine Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 02 | Feb 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 7595
WIND ENERGY STORAGE PREDICTION USING MACHINE LEARNING
S A K Jainulabudeen1, M.Sushil 2 ,Mohammed Afrith A 3,Rohith S4
1Assistant Professor+, Students2,3,4
Department of Computer Science and Engineering,
Panimalar Engineering College, Chennai, Tamilnadu, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Traditional (conventional) sources of power
generation are Thermal (Coal), Hydro, Gas and Nuclear, but
they are depleting and causing carbon emission. Many
countries are taking harsh decision to closedownthermaland
nuclear. An alternative source, catching the attention of
everyone is renewable energy from Solar and wind, which
require no fuel and abundantly available according to
geographical location. However, it has its own characteristic
and throw upon challenges to integrate into transmissionand
distribution grid. Though available throughouttheyear. Wind
potential varies location to location (that’s why installed in
specific areas mostly in remotelocations)andseasontoseason
Wind is variable, intermittent andunpredictable during24 hrs
of the day. Location wise wind potential makes the task of
transmission and distribution utility/grid operation more
difficult in absence of local consumption as well as adequate
network. Wind Energy is at peak during monsoon. This is the
season when power demand is low. Grid operation has a
challenge of handling the excess renewal energy. Grid
operation is planned day ahead by taking supply (generator)
and demand (utility) commitment. Grid operator is bound to
control supply-demand balance to maintain frequency, but it
becomes challenging whenwindenergyisaccountedaspartof
supply in the day ahead planning due to variable and
intermittent nature of wind. Grid operator is compelled to
back-down conventional sources of generation to minimum
level (inefficient operation) for load balancing.
Key Words Wind Energy ,Machine Learning ,Grids,
renewable energy, storage
1.INTRODUCTION
To solve climate change, society needs to
rapidly decarbonize Incorporation of greater amounts
of renewable wind powerwill be oneofmanyessential
strategies for decarbonization But the wind does not
always blow.
To incorporate high amountsofwindpower,thefuture
smart grid will require accurate forecasts to operate
effectively and efficiently.
Vertically-integrated electric utility company
seeking to incorporate higher amounts of wind power
generation into its electricity portfolio .employ
machine learning to accurately predict hourly wind
power generation at 7 wind farms, based on historical
wind speeds and wind directions.

Fig1. Wind Speed vs Wind Direction
Evaluation:
Root Mean Square Error (RMSE): a value of 0 indicates
a perfect fit. The lower the RMSE value, the more
accurate the predictive capability of the model
Solar energy refers to capturing the energy from the
Sun and subsequently converting it into electricity.We
can then use that electricity to light up our homes.
Data provided by Institute of Electrical and
Electronics Engineers (IEEE) .Retrieved from Kaggle
database.Time series dataset with wind speed, wind
direction and wind power production data.Data for 7
separate wind farms.Unitless for anonymity.
Fig 2. Wind Speed vs Wind Farm
Fig 3. Wind Speed Over the year
Fig 4. Wind farm
How is solar energy produced
• Solar energy refers to capturing the energy
from the Sun and subsequently converting it
into electricity.
• We can then use that electricity to light up our
homes.
2.RELATED WORKS
In their survey, Costa et al. [2] present an
expansive diagram of different techniques and
scientific, measurable and physical models utilized
over the most recent 30 years for shortterm
expectation. Soman et al. [19] give a broad overview of

the conceivable methods for various conjecture
skylines. Past outcomes have demonstrated that
techniques from factual learning are incredible
methodologies for transient vitality forecast. For
instance, Juban et al. [8] displayed a bit thickness
estimationapproachforaprobabilisticdeterminingfor
various breeze parks. Foresti et al. [5]utilizeddifferent
part learning relapse as an all-inclusive help vector
demonstrate that self-rulingly recognizes the
significant highlights for wind speed expectations.
Additionally neural systems have been
connected to wind control forecast before, e.g., by
Mohandes etal.[13],whocontrastedanautoregressive
model and a traditional backpropagation arrange. In
this line of research, Catalao et al. [1] prepared athree-
layered feedforward connect with the Levenberg-
Marquardt calculation for momentary breeze control
anticipating, which beat the constancy model and
ARIMA approaches.Further,Hanetal.[7]concentrated
on an outfit strategy for neural systems for wind
control forecast. As to total of wind turbines,
Focken et al. [4] considered the decline of the
forecast mistakeofatotaledpoweryieldbroughtabout
by spatial smoothingimpacts.Fromthepointofviewof
electrical specialists, Pöller and Achilles [16]
investigatedhowuniquebreezeturbinescanbetotaled
to a solitary generator. The spatio-transient breeze
control forecast methodology that is premise of our
line of research has been presented in [10] with an
increasingly broad portrayal in [11]. In [20], we
exhibited a methodology for preselection of turbines
for kNN-based expectation. As the enhancement issue
is hard to explain, we proposed a transformative Wind
Power Prediction with Machine Learning 15 blackbox
technique for a proficient element choice, which
compares to a determination of suitable turbines. In
[6], we proposed a gathering approach for SVR, where
little subsets of preparing information are arbitrarily
tested and the expectations of various SVRs are
consolidated to a solid classifier. As wind control
inclines are troublesome occasions for the joining into
the matrix, we considered this issue inadifferentwork
[12].
We treat incline expectation as arrangement
issue, which we tackle with SVMs. Recursive
component determinationdelineateshowthequantity
of neighbored turbines influences this methodology.
The issue of imbalanced preparing and test sets is
broke down concerning the quantity of no-incline
occasions. Practically speaking,sensorsmayfallflatfor
different reasons and the expectation models can't be
connected. In [17], we looked at different missing
information strategies for the attribution issue.
Another commitment of this work is a kNN-based
relapse technique, which is utilized as geo-attribution
preprocessing venture by considering the time
arrangement of the neighbored turbines. Last, in [22]
we broadened the collection of forecast techniques
with a cross-connection weighted k-closest neighbor
relapse (x-kNN) variation.
ThekNN-basedcomparabilitymeasureutilizes
loads that depend on the cross-connection of the time
arrangement of the neighboring turbines and the
objective. In the event that the cross-relationship
coefficient is high, the turbine gets a noteworthy
impact for the forecast by growing the comparing
measurement in the relapse display.

3.PROPOSED SYSTEM
1.Dataset Description
The data used for the project isprovidedbythe
Institute of ElectricalandElectronicsEngineers(IEEE),
Power & Energy Society, and retrieved through the
Kaggle database. The dataset is a time series dataset
with historical power generation, wind speeds and
wind directions, for the time period from July 2009 to
December 2010.
Data Cleaning Process and Reading the files
The raw dataset is comprised of several csv
files. The pandas library is imported as pd, and usedto
import the csvfilesusingpd.read_csv.Convertingdates
and times to DateTime objects
The dataset features a ‘date’ column, featuring
dates and times in integer (int64) format as
YYYYMMDDHH, where YYYY = year,MM=month,DD=
day, and HH = hour. To be more useful, these values
are converted from int64 into Date Time objects using
pd.to_datetime. Furthermore, these Date Time objects
are converted into a standardized ISO 8601 format for
convenience. A function called convert_to_iso is
defined, to convert a date time inint64formattoaDate
Time object in ISO 8601 format. The function
convert_to_iso is applied to the ‘date’ columns in the
dataset.
Creating a column for modified dates,
‘mod_date’ in addition to the ‘date’ column, the dataset
features an ‘hors’ column, featuring hour values in
int64 format. The values in ‘hors’ranges from 1 to 48,
representing the number of hours-ahead being
forecasted. For example, let’s say it is July 1st, 2009 at
12am. At this time and date, there is forecast data
associated with hors=1. This forecast data thus refers
to the forecast for the following time:
(July 1st, 2009, 12am) + (1-hour-ahead
forecast) = July 1st, 2009, 1am
Thus, a new column, ‘mod_date’, is created to
feature the DateTimes that include both the original
datetimes (‘date’), and the hour-ahead forecasts
(‘hors’).
‘date’ + ‘hors’ = ‘mod_date’
In the equation above, values for ‘date’ are in
DateTime format. Values for ‘hors’ are originally int64
but are converted to timedelta format using
pd.to_timedelta.
Creating a column for forecast categories,
‘forecast_cat’
Forecasting data is split into the following four
categories:
Category 1: 1-hour to 12-hour ahead data
Thus, a new column ‘forecast_cat’ is created,
featuring a value ranging from 1 to 4 for each datarow,
corresponding to the appropriate forecast category
shown above. Boolean selection is achieved using .loc.
Merging wind data with training data

Then, the wind data is merged with the wind
power generation training data using
pd.DataFrame.merge. The specific merge method is
‘left’, specified by the ‘how’ argument, andisanalogous
to a LEFT OUTER JOIN SQL Join. Thus, the left outer
join in this situation returns all of the rows for which
there is both wind speed, direction data and wind
power generation data available. Rows that havewind
power generation data but no wind speed, direction
data available are not included.
3.Machine learning Models
Predicting wind power production based on
wind speedandwinddirectionisaregressionproblem.
We modeladependentvariable(windpower)basedon
two independent variables (wind speed and wind
direction). In total, five machine learning modelswere
trained and tested on the data: Linear Regression,
Ridge Regression, Lasso Regression, Decision Tree
RegressionandNeuralNetworkRegression.RootMean
Square Error (RMSE) was the metric used to evaluate
prediction accuracy; an RMSE closer to a value of 0 is
indicative of higher predictive accuracy. Generally,
Decision Tree Regressors performed best, producing
the lowest RMSE values across all wind farms. Neural
Network Regressors performed worse. Also, Lasso
Regression and Ridge Regression produced virtually
the same results as standard Linear Regression. This
makes sense given thatthereareonlytwoindependent
variables at play; Lasso and Ridge are useful when
there are many more feature variables involved, and it
is necessary to incorporate variable selection.
Linear/Lasso/Ridge Regression generally did not
perform as well as Decision Tree Regression but
always outperformed Neural Networks.
Experimental results:
Fig 5 Wind Farm 1
Fig 6 Wind Farm 2
Fig 7 Wind Farm 3

Fig 8 Wind Farm 4
Fig 9 Wind Farm 5
Fig 10 Wind Farm 6
Fig 11 Wind Farm 7
4. CONCLUSIONS
A variety of Machine Learning Models were trained
and tested on wind speed, wind direction and wind power
production data. Prediction accuracy was evaluated using
Root Mean Square Error (RMSE). Decision Tree Regressors
performed best as compared to Standard Linear Regression,
Lasso Regression, Ridge Regression and Neural Networks.
Therefore, Decision Tree Regression is recommended for
future predictions of wind speed, wind direction and wind
power production data at the various wind farms.
REFERENCES
[1] Treiber, N.A et.al O.: Evolutionary turbine selection
for wind power predictions. In: 37th Annual German
Conference on AI, pp. 267–272 (2019).
[2] Soman, S.S., et.al.:A reviewof wind power and wind
speed forecasting methods with different time
horizons. In: North American Power Symposium
(NAPS),pp. 1–8 (2018).
[3] Lew, D., Milligan, et.al, R.: How do wind and solar
power affect grid operations: the western wind and
solar integration study. In:8th International
Workshop on Large Scale Integration of Wind
Power and on Transmission Networks for Offshore
Wind Farms (2018)
[4] Catalao, J.P.S., et.al V.M.F.: An artificial neural
network approach for short-term wind power
forecasting in Portugal. In: 15th International
Conference on Intelligent System Applications to
Power Systems (2009)
[5] Costa, A., et.al A review on the young history of
thewind power short-term prediction. Renew.
Sustain. Energy Rev. 12(6), 1725–1744 (2008)
[6] Ernst, B., et.al Predicting the wind. Power Energy
Mag. 5(6), 78–89 (2007)
[7] Focken, U., et.al A.: Short-term prediction of the
aggregated power output of wind farms—a
statistical analysis of the reduction of theprediction

error by spatial smoothing effects. J.Wind Eng. Ind.
Aerodyn. 90(3), 231–246 (2002)
[8] Foresti, L., et.al A.: Learning wind fields with
multiple kernels. Stoch. Env. Res.Risk Assess.25(1),
51–66 (2011)
[9] Heinermann, et.al Precise wind power prediction
with SVM ensemble regression.In: Artificial Neural
Networks and Machine Learning—ICANN 2014,pp.
797–804. Springer,Switzerland (2014)
[10] Han, S., et.al Neural network ensemble
method study for wind power prediction.In: Asia
Pacific Power and Energy Engineering Conference
(APPEEC) (2011)
[11] Juban, J., et.al Probabilistic short-term wind
power forecasting based on kernel. In: Density
Estimators. European Wind Energy Conference,pp.
683–688. IEEE (2007)
[12] Kramer, O et.al A framework for data
mining in wind power time series. In: Proceedings
of ECML Workshop DARE (2014)
[13] Kramer, O et.al Short-term wind energy
forecasting using support vector regression.In: 6th
International Conferenceon SoftComputingModels
in Industrial and Environmental Applications
(2011)
[14] Kramer, O., et.al Analysis of wind energy
time series with kernel methods and neural
networks. In: 7th International Conference on
Natural Computation (2011) .
[15] Kramer, O., et.al Wind power ramp event
prediction with support vector machines. In: 9th
International Conference on Hybrid Artificial
Intelligence Systems (2014)
[16] Mohandes, et.al A neural networks
approach for wind speed prediction.Renew.Energy
13(3), 345–354 (1998) .
[17] Pedregosa, F., et.al Scikit-learn: machine
learning in Python. J. Mach. Learn. Res. 12, 2825–
2830 (2011)
[18] Pöller, M., Achilles, S.: Aggregated wind
park models for analyzing power system
dynamics.In: 4th International WorkshoponLarge-
scale Integration of Wind Power and Transmission
Networks for Offshore Wind Farms, Billund (2003)
[19] Poloczek, J., et.al KNN regression as geo-
imputation method for spatiotemporal wind data.
In: 9th International Conference on Soft Computing
Models in Industrial and Environmental
Applications (2014) .
[20] Robusto, C.C.: The Cosine-Haversine
formula. Am. Math. Mon. 64(1), 38–40 (2018)
[21] . Treiber, N.A., et.al Aggregation of features
for wind energy prediction with support vector
regression and nearest neighbors. In: European
Conference on Machine Learning, DARE Workshop
(2017)
[22] Treiber, N.A., Kramer, O.: Wind power
prediction with cross-correlation weighted nearest
neighbors. In: 28th International Conference on
Informatics for Environmental Protection (2017).
[23] Wegley,H., et.al Subhourlywindforecasting
techniques for wind turbine operations. echnical
report, Pacific Northwest Lab., Richland,WA(USA).

IRJET- Wind Energy Storage Prediction using Machine Learning

More Related Content

What's hot (20)

Similar to IRJET- Wind Energy Storage Prediction using Machine Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Wind Energy Storage Prediction using Machine Learning