Optimize Short Term load Forcasting Anomalous Based Feed Forward Backpropagation

This paper contains the Short-Term Load Forecasting (STLF) using artificial neural network especially feed forward back propagation algorithm which is particularly optimized in order to getting a reduced error value result. Electrical load forecasting target is a holiday that hasn’t identical pattern and different from weekday’s pattern, in other words the pattern of holiday load is an anomalous. Under these conditions, the level of forecasting accuracy will be decrease. Hence we need a method that capable to reducing error value in anomalous load forecasting. Learning process of algorithm is supervised or controlled, then some parameters are arranged before performing computation process. Momentum constant a value is set at 0.8 which serve as a reference because it has the greatest converge tendency. Learning rate selection is made up to 2 decimal digits. In addition, hidden layer and input component are tested in several variation of number also. The test result leads to the conclusion that the number of hidden layer impact on the forecasting accuracy and test duration determined by the number of iterations when performing input data until it reaches the maximum of a parameter value.


Introduction
Load Forecasting becomes an important instrument in the power system operation, always. Many various operation decision are determined by load forecasting, such as generator capacity regulation schedule, test analysis, and generator maintenance plan. Besides that, load forecasting accuracy is an important key to predicting electric cost. Error forecasting significantly influence benefit, stock market, and holder value of course. [1].
Short term load forecasting approach which used in literature can devided become two general categories: statistic method and artificial intelligent based method. Statistic category is consist of multiple linear regression, stochastic time series, ARIMAX and general exponential smoothing, state space model, and support vector regression (SVR), meanwhile expert system, artificial neural network and fuzzy interference are include to artificial intelligent category [2].
Artificial intelligent is very booming in this last two decades. This compute method application very interested and applied for many kind problem. Artificial intelligent gives strong and flexible reason to decide solution of various problem that frequency unsolved by traditional anf orthodox method [3]. Because forecast result proved better accuracy than traditional method. Gradien reduce algorithm such as backpropagation (BP) or the variation in multi-layer feedforward network has used for many aplication extensively. But, the most serious problem about BP is local minima. Then used algorithm modified in order to be stable result althougt using many hidden layers. [4].

1)
Training data process: backpropagation network process involve three stages that consist of input training feedforward pattern, link error backpropagation, and weight adjustment.
Feed forward backpropagation algorithm as follow : a) Step 0: Weight insialisation (set to the random small value). b) Step 1: When stop condition not suit, do step 2-9. c) Step 2: For training data, do step 3-8. Feed forward : d) Step 3: Each input unit (Xi, i = 1, . . . , n) receives input signal xi and spreads to all hidden layer units. e) Step 4: Each hidden unit (Zj, j = 1, . . . , p) collect input weight signal, (iv) Step 9: Test stop condition. [5] [6]  Output learning result is an forecast electric load which is needed to compute error accuracy.
Repeat data train process as many as 5 times.

Result and Analysis
Indonesia using different load electric every time. On weekday, electric load disposed stable. Weekend is different from weekday, where electric load consume can change suddenly dependent to event society. Fig. 1 below shows how the comparison between load electric consume of weekday and national weekend with take 16 days as sample.

Load Forecasting using Feed Forward Backpropagation Algorithm
Feed forward backpropagation algorithm has probabilistic and random characters, thus with a same epoch value obtained different result always and derived the best average result. Optimum result will derived with ten time performs [6].

Figure 2. Comparison Output Graph Learning Result from Target
Learning result error percentage value is 2,27% suit to used as anomalous load forecasting. However, the graph above show not careful target pattern especially on the peak load. Forecast result explain that plot keep away from 5.30 p.m to 10.30 p.m with maximum accomplishment 3694,2 MW meanwhile the peak load of target on that time accomplish 3947,58 MW.

Optimation of Short Term Load Forecasting for Anomalous Load A subsubsection.
Learning mistake reduced with optimation. Hereafter, anomalous electric load forecasting research result turn out to be more careful and increase accuracy value. Although each trial result is different, but forecast output has expected to approaching target even if data is anomalous load. When that happen, thus accuracy value of feed forward backpropagation learning will be higher.   If each method is compared by target, then load coefficient has higher error forecast that 5,46 therefore the accuracy become 94,54%. This occasion will impact graph formation which is shown in  Research toward two different methods give a conclusion that feed forward backpropagation method proved more effective than conventional method such as load coefficient to perform short term load forecasting on anomalous load case because accuracy of load coefficient cannot suit as good category yet.

Conclusion
Load consume pattern in Indonesia's holiday is different with weekday. Historical data of PT. PLN (Persero) APB Region II West Java explain electric load consume on holiday much lower than weekday and load peak unexpected occur. Great days exclusively Idul Fitri caused anomalous load pattern formed.
Feed forward backpropagation algorithm suit to conquer STLF problem of anomalous load. This algorithm forecast accuracy established very effective within very small error accuracy. Besides that, this algorithm excellence is hidden layer generate accuracy extensively. But also this algorithm is random and probabilistic and then some trials required for each parameter.
Optimation process perform toward hidden layer and input component, resumed to epoch and learning rate parameter. The result is more odder number of hidden layer will generate accuracy significantly and each layer compute duration disposed to be same. Different of maximal epoch is not much influence toward error accuracy and duration. Each learning rate value take various duration, careful excalating from learning rate not impact to accuracy result and duration. Outfit learning rate value for feed forward backpropagation algorithm will create the best error accuracy. Number of input perform linear with its accuracy result. Compute duration reliant on number of iteration that necessary.