Short-Term Load Forecasting of Australian National Electricity Market by Hierarchical Extreme Learning Machine
Download as PPTX, PDF0 likes321 views
This document summarizes a study on short-term load forecasting of the Australian National Electricity Market using a hierarchical extreme learning machine model. It first introduces short-term load forecasting and the benefits of the hierarchical extreme learning machine approach. It then describes the implemented hierarchical extreme learning machine structure, including the input data used, pre-processing strategies, and model architecture. The results show that the hierarchical extreme learning machine model achieved more accurate predictions with greater stability compared to the basic extreme learning machine model. In conclusion, the hierarchical extreme learning machine is an effective approach for short-term load forecasting when combined with appropriate data pre-processing.
![1. Introduction
• Predicting future demand load for electricity is considered an
essential process in power system operation.
• Good quality Short-Term Load Forecasting (STLF) is very
important for economical aspect and contributes to reliability of
power system [5], [6].
• There have been continuous efforts to achieve high load
forecasting accuracy.
• Improvement of the performance of the learning algorithm model
• How well the input data features are extracted through the pre-
processing process](https://image.slidesharecdn.com/elm2017-170807050211/85/Short-Term-Load-Forecasting-of-Australian-National-Electricity-Market-by-Hierarchical-Extreme-Learning-Machine-3-320.jpg)
![2.a. Extreme Learning Machine theory
• Pros:
• Due to the characteristics of the ELM based SLFN, we can obtain very
short training time, excellent efficiency and good generalize
performance [3].
• Cons:
• It randomly selects the input weights and biases for hidden nodes, and
cause a crux in the stability of its outputs [4].](https://image.slidesharecdn.com/elm2017-170807050211/85/Short-Term-Load-Forecasting-of-Australian-National-Electricity-Market-by-Hierarchical-Extreme-Learning-Machine-7-320.jpg)
![3.c. Correlation study
• NEM load data will be used not only as label data but also as
feature data
• Selecting the right amount of data and the appropriate features of data as
input data is an important process that must be performed first for accurate
load forecasting [15].
• We used Pearson's correlation coefficient (PCC) to track the
relationship between date and load data.](https://image.slidesharecdn.com/elm2017-170807050211/85/Short-Term-Load-Forecasting-of-Australian-National-Electricity-Market-by-Hierarchical-Extreme-Learning-Machine-14-320.jpg)
![6. References
10. Jiexiong Tang, Chenwei Deng, Guang-Bin Huang, "Extreme Learning Machine for Multilayer Perceptron," IEEE transactions on
neural Network and learning systems, Vol. 27, No. 4, April 2016.
11. Pradeep C.Gupta, Keigo Yamada, "Adaptive Short-Term Forecasting of Hourly Loads Using Weather Information," IEEE
Transactions on Power Apparatus and Systems, Volume: PAS-91, Issue: 5, Sept. 1972, pp. 2085-2094
12. Muhammad Qamar Raza, Zuhairi Baharudin, Badar-Ul-Islam, Mohd. Azman Za- kariya, Mohd Haris Md Khir, "Neural Network
Based STLF Model to Study the Seasonal Impact of Weather and Exogenous Variables," Intelligent and Advanced Systems (ICIAS),
2014 5th International Conference on, August 2014.
13. S.J.Kiartzis, C.E. Zoumas, A.G.Bakirtzis,V. Petridis, "Data Pre-Processing For Short-Term Load Forecasting In An Autonomous
Power System Using Artificial Neural Networks," Electronics, Circuits, and Systems, 1996. ICECS ’96., Proceed- ings of the Third
IEEE International Conference on, October 1996
14. Agnaldo J. Rocha Reis, Alexandre P. Alves da Silva, "Feature Extraction via Mul- tiresolution Analysis for Short-Term Load
Forecasting," IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 1, February 2005.
15. Abdussalam Mohamed, M. E. El-Hawary, "Effective Input Features Selection For Electricity Price Forecasting," 2016 IEEE
Canadian Conference on Electrical and Computer Engineering (CCECE)
16. Bunn, D.W., Farmer, E.D. (Eds.), "Comparative models for electrical load fore- casting" (John Wiley & Sons, 1985)
17. Australian Energy Market Operator [Online]. Available at http://www.aemo.com. au
18. Bureau of Meteorology of Australian Government [Online]. Available at http:// www.bom.gov.au/index.shtml](https://image.slidesharecdn.com/elm2017-170807050211/85/Short-Term-Load-Forecasting-of-Australian-National-Electricity-Market-by-Hierarchical-Extreme-Learning-Machine-25-320.jpg)
Short-Term Load Forecasting of Australian National Electricity Market by Hierarchical Extreme Learning Machine
- 1. Short-Term Load Forecasting of Australian National Electricity Market by Hierarchical Extreme Learning Machine Park Jee Hyun 07 JUL 2017 - SungKyunGwan University – undergraduate student Nanyang Technological University – exchange student
- 2. Contents 1. Introduction 2. Hierarchical extreme learning machine (H-ELM) 3. Implemented structure 4. Test results 5. Conclusion 6. References
- 3. 1. Introduction • Predicting future demand load for electricity is considered an essential process in power system operation. • Good quality Short-Term Load Forecasting (STLF) is very important for economical aspect and contributes to reliability of power system [5], [6]. • There have been continuous efforts to achieve high load forecasting accuracy. • Improvement of the performance of the learning algorithm model • How well the input data features are extracted through the pre- processing process
- 4. 2. Hierarchical Extreme Learning Machine a. Extreme Learning Machine theory b. H-ELM framework
- 5. 2.a. Extreme Learning Machine theory
- 6. 2.a. Extreme Learning Machine theory • ELM learning can be summarized as the following major three steps:
- 7. 2.a. Extreme Learning Machine theory • Pros: • Due to the characteristics of the ELM based SLFN, we can obtain very short training time, excellent efficiency and good generalize performance [3]. • Cons: • It randomly selects the input weights and biases for hidden nodes, and cause a crux in the stability of its outputs [4].
- 8. 2.b. H-ELM framework • HELM consists of two parts 1) Unsupervised hierarchical feature representation 2) Supervised feature classification
- 9. 2.b. H-ELM framework 1) Unsupervised hierarchical feature representation • It receives raw input data and projects it to the ELM random feature space, which plays an important role in extracting hidden feature information of input training data. • High-level sparse features can be obtained as output values through N- layer unsupervised learning, which is then passed to the input data of the second stage. 2) Supervised feature classification • The final decision is made through regression, which is the same process as the existing ELM.
- 10. 3. Implementation Structure a. Data used in experiments b. Data pre-processing strategies c. Correlation study d. STLF architecture e. Input & Output
- 11. 3.a. Data used in experiments • Australian National Electricity Market (NEM) • Contents : • Load profile of New South Wales (NSW) region from 1 January 2006 to 31 December 2015 • k-th date of the data : • Australian Bureau of Meteorology (BOM) • Contents : • ㅍof NSW region from 1 January 2006 to 31 December 2015 • the k-th date of the data :
- 12. 3.b. Data pre-processing strategies • The feature of the weather data is too poor to be used as the input feature data of the STLF. • BOM data show only one record per day for maximum temperature, minimum temperature, rainfall, and solar exposure. • Using these data directly as input feature dataset can not provide sufficient features to train the STLF model • To overcome this problem, we suggest following data pre- processing strategy • To enrich the features of input data, include the historical load data from the NEM in the input dataset consisting of the weather data of the BOM.
- 13. 3.b. Data pre-processing strategies • In short, the forecasting method is based on historical load data and takes weather forecast as a hint to predict future load data.
- 14. 3.c. Correlation study • NEM load data will be used not only as label data but also as feature data • Selecting the right amount of data and the appropriate features of data as input data is an important process that must be performed first for accurate load forecasting [15]. • We used Pearson's correlation coefficient (PCC) to track the relationship between date and load data.
- 15. 3.c. Correlation study • As a result, the load data of the previous day and the week before showed the highest PCC value. • We set the strategy to add the load data of just before and one week before to input feature dataset.
- 16. 3.d. STLF architecture • We will compare the performance of each STLF architecture implemented with ELM and H-ELM.
- 17. 3.e. Input & Output • Input data consists of feature data and label data • Output value is forecasted daily load profile, namely the 48 load points corresponding to each half-hour of a day.
- 18. 4. Test Results a. Optimal number of hidden nodes b. Performance comparison between ELM and H-ELM
- 19. 4.a. Optimal number of hidden nodes • The STLF model designed in this experiment needs to be optimized for maximum performance, which is the task of determining the number of hidden nodes.
- 20. 4.b. Performance : ELM vs. H-ELM • The accuracy of the prediction is evaluated using MAPE and MAE. • The stability of the prediction is evaluated as the standard deviation of the distribution of the prediction accuracy. ➔In order to measure these two evaluation items, 1,000 identical repeating experiments were conducted for each STLF model. ➔In order to evaluate the accuracy, the mean value of MAPE and MAE values was taken as the final value, and the standard deviation was calculated to evaluate the stability.
- 21. 4.b. Performance : ELM vs. H-ELM • Forecasting MAPE and stability
- 22. 4.b. Performance : ELM vs. H-ELM • Forecasting MAE and stability
- 23. 5. Conclusion • Improvement of Learning Algorithm • Due to the characteristics of ELM, the output value of the ELM SLTF is unstable, which degrades the accuracy of prediction. • By constructing the STLF model through H-ELM, it is possible to solve the instability of the output value and able to get more accurate prediction. • Data pre-processing • Data pre-processing plays a significant role in improving the performance of the STLF model. ➔ In summary, in order to improve the performance of the STLF, it is necessary to construct a STLF model using H-ELM and to input well-defined input features through proper data pre- processing.
- 24. 6. References 1. Zhiyi Li, Xuan Liu, Liyuan Chen, "Load Interval Forecasting Methods Based on An Ensemble of Extreme Learning Machines," Power & Energy Society General Meeting, 2015 IEEE, October 2015. 2. Guang-BinHuang,ErikCambria,“ExtremeLearningMachines,”IEEETransactions on Cybernetics, vol. 28, no. 6, pp. 30-59, 2013. 3. G.-B.Huang,Q.-Y.Zhu,andC.-K.Siew,“ExtremeLearningMachine:ANewLearn- ing Scheme of Feedforward Neural Networks,” 2004 International Joint Conference on Neural Networks (IJCNN’2004), July 25-29, 2004. 4. Rui Zhang, Zhao Yang Dong, Yan Xu, Ke Meng, Kit Po Wong, "Short-term load forecasting of Australian National Electricity Market by an ensemble model of ex- treme learning machine," December 2012. 5. Muhammad Qamar Raza, Dr-Badar Islam, M. A. Zakariya, "Neural Network Based STLF Model to Study the Seasonal Impact of Weather and Exogenous Variables," Research Journal of Applied Sciences, Engineering and Technology · November 2013. 6. Damitha K. Ranaweera, George G. Karady, Richard. C. Farmer, "Economic Impact Analysis of Load Forecasting," IEEE Transactions on Power Systems, Vol. 12, No. 3, August 1997. 7. C.L. Wu, K.W. Chau, C. Fan, "Prediction of rainfall time series using modular artificial neural networks coupled with data- preprocessing techniques," Journal of Hydrology Volume 389, Issues 1–2, 28 July 2010. 8. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K, "Extreme learning machine: theory and ap- plications," Neurocomputing, 2006, 70, pp. 489–501 9. Bartlett, P.L., "The sample complexity of pattern classification with neural net- works: the size of the weights is more important than the size of the network," IEEE Trans. Inf. Theory, 1998, 44, (2), pp. 525–536
- 25. 6. References 10. Jiexiong Tang, Chenwei Deng, Guang-Bin Huang, "Extreme Learning Machine for Multilayer Perceptron," IEEE transactions on neural Network and learning systems, Vol. 27, No. 4, April 2016. 11. Pradeep C.Gupta, Keigo Yamada, "Adaptive Short-Term Forecasting of Hourly Loads Using Weather Information," IEEE Transactions on Power Apparatus and Systems, Volume: PAS-91, Issue: 5, Sept. 1972, pp. 2085-2094 12. Muhammad Qamar Raza, Zuhairi Baharudin, Badar-Ul-Islam, Mohd. Azman Za- kariya, Mohd Haris Md Khir, "Neural Network Based STLF Model to Study the Seasonal Impact of Weather and Exogenous Variables," Intelligent and Advanced Systems (ICIAS), 2014 5th International Conference on, August 2014. 13. S.J.Kiartzis, C.E. Zoumas, A.G.Bakirtzis,V. Petridis, "Data Pre-Processing For Short-Term Load Forecasting In An Autonomous Power System Using Artificial Neural Networks," Electronics, Circuits, and Systems, 1996. ICECS ’96., Proceed- ings of the Third IEEE International Conference on, October 1996 14. Agnaldo J. Rocha Reis, Alexandre P. Alves da Silva, "Feature Extraction via Mul- tiresolution Analysis for Short-Term Load Forecasting," IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 1, February 2005. 15. Abdussalam Mohamed, M. E. El-Hawary, "Effective Input Features Selection For Electricity Price Forecasting," 2016 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE) 16. Bunn, D.W., Farmer, E.D. (Eds.), "Comparative models for electrical load fore- casting" (John Wiley & Sons, 1985) 17. Australian Energy Market Operator [Online]. Available at http://www.aemo.com. au 18. Bureau of Meteorology of Australian Government [Online]. Available at http:// www.bom.gov.au/index.shtml
Editor's Notes
- #6: Given a training data set with N samples, the output function of the SLFN with L hidden nodes and activation function θ. The activation function is expressed as (1).
- #7: ELM is completely different from traditional iterative learning algorithms as it randomly selects the input weights and biases for hidden nodes, ω and b and analytically calculates the output weights, β, by finding least-square solution [8]. once ELM parameters of hidden layers of the ELM are generated randomly, it does not need to be tuned. Therefore, its training speed can be thousands of times faster [3].
- #8: Due to the characteristics of the ELM based SLFN, we can obtain very short training time, excellent efficiency and good generalize performance [3]. However, it randomly selects the input weights and biases for hidden nodes, and cause a crux in the stability of its outputs [4].
- #10: Since H-ELM receives input of high-level sparse features obtained from unsupervised feature learning, instead of raw input data, it is able to obtain more accurate and stable performance. The experiments designed in the back will confirm that H-ELM shows better performance than ELM.
- #12: Normally in STLF, load data is used as input label data and weather data is used as input feature data.
- #13: Which historical load data to be included in the feature input dataset will be described in the following correlation study.
- #17: This architecture is divided into two phases: a) training and b) testing. In the a) training phase, the learning model receives the feature vector obtained through the data pre- processing process and performs model training. The STLF model trained in this phase is used in the next phase. In the b) testing phase, the feature vector is input and forecasting is performed based on the STLF model trained in the previous process.
- #18: In this experiment, we will confirm the role of data pre-processing in addition to the performance comparison between ELM and H-ELM. Therefore, based on the PCC results, we propose the following 4 cases.
- #20: The STLF model designed in this experiment needs to be optimized for maxi- mum performance, which is the task of determining the number of hidden nodes. the MAPE method was used to measure accuracy, and the forecasting accuracy result obtained from this optimization process corresponds to the average value of the same 20 repetitive prediction experiments. In each case, the number of hidden nodes for the optimization of the STLF model slightly differed, and the results of the experiment for the optimization are shown in Figure.