presentation: wind speed prediction using radial basis function neural network
TRANSCRIPT
Arzam Muzaffar Kotriwala UNMKL_009994
Wind Speed Prediction Using Radial Basis Function Neural
Network
H53PJ3 Final Year Individual Project
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Motivation | Why Predict Wind? ² Increase in demand for renewable energy
• Increase in crude oil prices
• Worldwide awareness of environmental issues & energy scarcity
² Wind power characteristics
• Environment-friendly
• High efficiency
² Power production capacity varies greatly with varying weather conditions
² Short term predictions are useful for:
• Administering wind power
• Scheduling maintenance
• Boosting power generation efficiency
² To prepare for anticipated destruction caused by high speed winds and
catastrophes such as hurricanes
Motivation | Why Neural Networks?
² Wind exhibits non-linear behavior.
² Neural networks are capable of handling non-linear data.
² A simple approach for solving various problems that are otherwise difficult
to be modeled by conventional methods
² Neural network have the ability to:
• Learn from data/examples
• Recognize conspicuous and hidden patterns in chronological
observations
• Use these relationships to predict forthcoming data
² Suitable for wind speed prediction owing to:
• Simplicity
• Robustness
Motivation | Applications of Neural Networks? ² Pattern Recognition
² Optimization
² Power Systems
² Medicine
² Robotics
² Control Systems
² Manufacturing
² Signal Processing
² Psychology
² Forecasting
• Weather and market trends
• Predicting mineral exploration sites
• Electrical & Thermal load predictions
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Objectives Deliverables
Obtain and organize historical wind data to train network
Variables relevant to wind speed prediction identified and separated into distinct training and prediction sets
Design a RBF neural network with appropriate input parameters and architectures
Short-term wind speed forecasting models with various network configurations
Develop code to implement and train the neural network models
Validation of forecasting accuracy of RBF model with plots and error calculations
Test the performance to investigate and analyze the RBF prediction technique
Justifications for using the proposed RBF neural network design
Project Objectives & Deliverables
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Biological Neural Networks
Learning in biological structures entails modifications to the synaptic
weight connections that link the neurons.
Artificial Neural Networks What are they?
² Inspired by the biological neural system
² A subset of the domain of AI
How do they work?
² Each single neuron is connected to other neurons of a previous layer
through adaptable synaptic weights
² Patterns are stored as a set of connection weights
Radial Basis Function Neural Network
² Activation function = Radial Basis Gaussian function
² RBF has been previously applied across a spectrum of engineering
problems.
² Studies have revealed that the BP network converges at a slow pace.
² RBF supersedes BP in terms of learning speed and approximation accuracy
and is free from the local minima problems of BP models.
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Methodology
The RBF wind speed forecasting system is summarized as below:
Historical Data Collection
Data Assimilation
Prediction Using RBF
Comparison with Other Techniques
Methodology | Historical Data Collection ² The data is obtained from the Weather Analytics database and is used to
train, test and validate the network.
² The data comprises wind speed values recorded hourly in knots, measured
at the Weather Analytics meteorological station in Madrid, Spain.
² The wind power time series data was recorded for one complete year, from
January 1, 2012 to December 31, 2012.
² The choice of training data plays a significant role in the overall performance
and training convergence of the neural network models.
² Division of data:
• Training data = 250 days
• Testing data = 106 days
Methodology | Data Assimilation
² Linear and multiple regression used to isolate the most important
independent variables.
² Number of inputs to neural network:
• Autocorrelation
• Partial Autocorrelation
² Normalization of data
Methodology | Prediction
² The neural network model uses real world historical hourly wind data as
examples to learn from.
² Upon the presentation of each training example, the network produces an
output based on the input pattern, which is then compared with the correct
desired output of the training pattern. In the case of there existing a
difference in these two values, the synaptic weights are changed in a
direction such that that the error is reduced.
² Once trained, the RBF model is expected to perform projections and
generalizations at high speed.
² Design parameters
• Number of hidden neurons
• Activation function
• Spread factor
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Results | Performance
Error Metrics:
² Root mean square error (RMSE)
² Mean absolute error (MAE)
² Mean absolute percentage error (MAPE)
The performance of the RBF neural network model is compared with:
² The Persistence theorem
² Back Propagation (BP) MLP neural network
Results | Summary of RBF Models
Model # of Inputs Spread Factor Hidden Neurons
RMS Error
A 4 10 30 1.69
B 8 50 30 1.70
C 8 14 80 1.64
D 12 90 43 1.65
E 4 10 8 0.56
Results | Summary of RBF Models
Comparison of the expected one-hour ahead output with the actual output of the
neural network of the response of Model C.
Results | Comparison with Persistence
Both the neural networks significantly outperformed the persistence technique.
0
0.5
1
1.5
2
2.5
3
3.5
4
0 1 2 3 4 5 6 7 8 9 10
RM
S Er
ror
Look Ahead Hours
RBF (VS) Persistence
RBF Persistence
Results | Comparison with MLP
Different initial connection weights of the MLP result in different training and
prediction performances.
Though the accuracies of the two networks differed slightly, the RBF model
proved to be more reliable and suitable for the task at hand.
Agenda
1. Motivation
2. Objectives & Deliverables
3. Project Fundamentals
4. Methodology
5. Results
6. Conclusion
Conclusion
² The results confirm the outcomes achieved by other researchers - the
applicability of neural networks to wind speed prediction is affirmed.
² Artificial neural networks are a reliable method for prediction.
² It was discovered that, different network configurations directly influenced
the forecast accuracy.
² It should be noted that the optimal choice of neural network or error metric
for a specific site may not necessarily be the most suitable option for
another site.
Conclusion
² Both, RBF and MLP neural networks predicted the time series fairly well.
However, certain consistent trends were seen in the errors. This could mean
that the neural networks are unable to predict the series to a high degree of
precision.
² The Radial Basis Function (RBF) network is advantageous over the Back-
propagation (BP) network in terms of consistency and reliability. Given a set
of training inputs and corresponding targets, the RBF network produced the
same result each time.
² Predicting the Beaufort Force of the wind revealed the usefulness of using
neural networks in wind speed prediction.
Recommendations for Future Work
² Weather is a continuous, multi-dimensional, data-intensive, dynamic and
chaotic process.
² Owing to these characteristics, highly accurate weather forecasting remains
a big challenge.
² Improvements to proposed models
• Training the network with data of more number of years.
• Reducing complexity of the designs – reducing number of hidden
neurons.
² Development of a single universal performance score – combining metrics
such as RMSE, MAE, MSE.