Residential Load Forecasting Usin

0 downloads 0 Views 173KB Size Report
LSTM networks. 1. Introduction. Residential load forecasting is still challenging task due to its complex and volatile phenomena. As consumed load happens in ...
2018년도 대한전기학회 전력경제연구회 춘계학술대회 논문집 (2018. 5.3-4)

Res i dential Load

For ec as ting

Us i ng CNN-LSTM Deep

Lear ning

Netw ork s

Shree Krishna Acharya and Jaehee Lee Mokpo National University

Abstract

- The occurrence of the residential loads are highly complex, volatile and non-linear. However, adequate prediction results facilitate the proper demand response over number of households. In this paper, a novel residential household’s load forecasting is proposed using deep learning methods. The proposed method uses the combination of convolution neural networks (CNN) and long-short term memory (LSTM) networks. With the aid of CNN and LSTM performing ability, the purposed method shows better performance measure as compared with CNN and LSTM networks.

1. Introduction

Residential load forecasting is still challenging task due to its complex and volatile phenomena. As consumed load happens in periodic time, multiple time series load forecasting method is useful measure to facilitate demand response and energy efficiency programs. Meanwhile, MTS load forecasting is attached with deep learning network (DNN) due to availability of huge amount of recorded load consumed values. The measured values are easily available using advanced metering infrastructure (AMI). Since extensive research has been done in statistical f orecasting methods such as moving average (MA), au toregressive (AR), ARMA, and ARIMA, some researc hers are still confused about the implementation of D NN in household’s load forecasting. However, deep for ecasting is the major concern for many subject areas like, price forecasting, energy forecasting, demand fore casting. The state-of-art deep learning network includ e back-propagation neural network (BPNN) convolutio n neural network (CNN), and recurrent neural networ k (RNN) [1 ]. Long-short memory (LSTM) is a specia l type of RNN which shows better forecasting perfor mance. In this paper, short-term residential load forec asting has been carried-out by using combination CN N and RNN. CNN is able to recognize the short cha nges of load series whereas RNN can capture the lon g changes. However, input time series should contain the repetitive patterns for accurate result. The major l everage of household’s periodic load profiles is their s -

-

patial curve shows quite repetitive pattern [2 ]. The oc currence of repetitive load profiles occurs either in dai ly or weekly periods. 2. DNN Background

CNN and LSTM are the recently evolved deep learn ing networks particularly for images classification and speech translation. However, they showed significant performances while implementing the forecasting task. CNN is capable to detect the shift-invariant pattern w hich is known as short term dependencies [3 ]. Similarl y, LSTM is able to capture long term variation from the repetitive pattern of sequence load series [4 ]. So f ar, many researches are conducted using CNN and LS TM separately. CNN can generate very fast result wh ile it contains ReLU activation functions. In contrast, LSTM contained tanh and sigmoid activation function s inherently, that makes its computation time is slow. However, if the sequences contain some repetitive patt erns, LSTM shows higher performance measure amon g the deep learning networks. While performing house hold’s load forecasting using DNN, selection of activat ion function is crucial. 3. Purposed Forecasting Methods

The advanced metering infrastructure (AMI) able to record the each household’s periodic energy consumption in time series as               ∈  where one denotes 1-D and  is the length of load curves for training process. To do forecasting task,  is converted to the multiple time series as  ×                   ∈  where each  include m-step’s shapes of load curves and m is m-step lookback horizon of time series [5]. When  is applied to the CNN, it recognized  has width      , height  and depth is one. While training the load series, CNN required  number of filter with  ×  size. The convolution between  and  makes the dimension of transformed  ′ is large.  ′ is minimized using max-pooling network but it preserves the inherent property of load series. The proposed DNN leaning network’s can be demonstrate in Figure 1. By default, LSTM networks possess memory, where it stored the contemporary correlation relationship at current time step t.Due to auto-correlation relationship of load time-series,

LSTM network can recognize the load pattern while performing training process.

Proposed CNN-LSTM deep learning forecasting for residential households. Fig.1.

(a)

Fully connected network is attached directly after LSTM hidden layers which is crucial for any kinds of forecasting task. Although LSTM uses much time for training process due to sigmoid activation function, CNN can help to reduce computational time of overall proposed deep learning network. The performance of CNN can be enhanced by using ReLU activation function. Each CNN layer with activation function can be expressed as          max     

(1)

where      If    gradient becomes constant and if  ≤  large number of neural unit can facilitate accurate prediction of the load series. 4. Test Results

The proposed method is tested using one household’s 30 days’ hourly consumed energy values. The multiple time series contains one-day is obtained by taking on e-day look-back over sequence form of load series. The training process, validation process and testing p rocess of DNN used 65%, 20% and 15% of load series , respectively. The proposed method contains one inpu t and output layer with three hidden layers. The hidd en layer consists of one CNN, one LSTM and one ful ly connected layer. The input layer also used CNN pa rameters. To evaluate the forecasting performance of proposed method, LSTM and CNN are taken as a refe rence. The CNN uses twenty-four filters in input laye rs and twelve filters in the first hidden layer. The ma x-pooling takes matrix that reduce the large data set by nine-times with the maximum values. Root mean square (RMSE) and mean absolute percentage error (MAPE) are powerful comparison tools used to evalua te the four different types. The combination of CNN a nd LSTM produced RMSE within a range of 0.163 to 0.175 which is lower than compared to CNN and LST M as shown Fig. 2. (a). Similarly, The MAPE results of the CNN-LSTM network is satisfactory with aroun d 30%. Due to small amount of data is used for traini ng process, overfitting is the main cause for lower M APE result. -

-

Fig.2 .

(b)

Comparison of performance measure (a) RMSE (b) MAPE. 5. Conclusion

The combination of the CNN and LSTM facilitate the more accurate forecasting performance with adequate t ime. The training time can be controlled using numbe r of layers designed by CNN and LSTM. The perfor mance measure lead to research about high amount of neighborhood data. [Acknowledgement]

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government(MSIT) (No. 2017R1C1B5016244). This work is financially supported by the Ministry Of Trade, Industry & ENERGY(MOTIE) through the fostering project of Energy valley industry-academy cooperation district furtherance business [Reference]

K. Amarasinghe, D. L. Marino, M. Manic,” Deep Neural Network for Energy Load Forecasting”, IEEE 26th Internatio nal Symposium on Industrial electronics (ISIE), 19-21 June 2 017. [2] B. Stephen, X. Tang, P. R. Harvey, S. Gall0woy, K.I. Je nnett, “Incorporating Practice Theory in sub-profile Models f or Short-Term Aggregated Residential Load Forecasting”, IE EE Transactions on Smart Grid, Vol.8, Issue no. 4, pp.15911598, 09 September 2017. [3] G. Lai, W. Chang, Y. Yang, H. Liu,”Modelling Long –an d Short-Term Temporal Patterns with Deep Neural Network s”, Cornell University Library, https://arxiv.org/abs/1703.0701 [1]

5.

W. Kong, Z.Y. Dong, Y. Jia, D.J. Hill, Y. Xu,” Short-ter m residential load forecasting based on LSTM recurrent neu ral network,” IEEE Transactions on Smart Grid, 18 Septemb er 2017. [5] S. K. Acharya, S. Park, and J. Lee, “Short-Term Solar Power Forecasting Using Sequential Deep Learning Method,” KIEE Conference, Nov., 2017. [4]

-

-