Application of Deep Learning LSTM in Online Power Prediction on Three-Phase Power Transformer

–Electrical energy plays an important role in daily life, especially companies. The 3 Phase Power Transformer is one of the important electrical components that is very influential in distributing electrical energy to companies as is the case in PT. Semen Baturaja (Persero). 3-phase power transformers require attention because they are one of the components that are prone to interference, this interference can hinder the effectiveness of using electrical energy as a company support for employees to work. As one of the disturbances for 3-phase power transformers is overload or excessive power usage, overload can raise the temperature at the winding and reduce its service life. Artificial intelligence can be one of the keys to predict the use of power transformers in the future, especially deep learning by utilizing the LSTM algorithm. Optimal power prediction requires a lot of maximum input variables so that in this study, it not only adds an offline learning mode but adds a learning mode that can directly access the company's Power Quality Monitoring (PQM) website online with an average accuracy value of 86.62%.


Introduction
The 3-phase power transformer is an important asset of the power grid and monitoring its operating conditions is essential for functional and economic reasons. Regular power monitoring to ensure failures in 3-phase power transformers in the early stages is very important as a protection system because 3-phase power transformers are highly susceptible to interference. Especially overloaded power, in addition to being able to increase the temperature that is in the isolation of a 3-phase power transformer, power overload can reduce the service life of the power transformer. [1][2] [3][4] [5] 3-phase power transformer in PT. Semen Baturaja (Persero) Tbk Kertapati site has a capacity of 1600KVA with an input voltage of 6KV and an output voltage of 400 volts.
The development of technology plays an important role in all aspects of work. According to Jahromi and R. Piercy in 2019, artificial intelligence technology itself has been widely applied in everyday life such as smart homes, decision support systems, and sensor readings. Without realizing it, information technology is one of the main needs in business development, especially electrical energy that supports the business itself. [6][7][8] [9] Machine learning itself is a branch of artificial intelligence that is useful for predicting power in the future. Machine learning is a branch of evolving computational algorithms designed to mimic human intelligence by learning from the surrounding environment. They are considered working horses in a new era called big data. [10][11] [12] In short, machine learning is a field of science that enables a computer program to learn from a set of data A.A. Ningrum, et al. (A 2021) were able to conduct an analysis to predict the age or feasibility of transformers using the Long Short Term Memory (LSTM) algorithm, but the result of the discussion was a transformer that had not been developed from an IoT perspective, which was the era, and could not be predicted in real time.
LSTM is a Deep Learning algorithm, which is a derivative of machine learning. The LSTM algorithm can be used to predict the use of power consumption to protect the transformer from excess power otherwise known as overload power. [6] [13] Data on Power Quality Monitoring (PQM) in PT. 3-phase power transformer. Semen Baturaja (Persero) Tbk will be inputted and labeled then the data will be trained. The results of the training data will produce a prediction which is then tested first for its suitability with the desired results, if appropriate, it will produce outputs and outputs need to be retested and adjusted through test data so that the output results are more accurate. [14] [15] [16]

LSTM (Long Short Term Memory)
A Long Short Term Memory Network (LSTM) is a modified version of a recurrent neural network or RNN. There are many changes to the RNN, but LSTM is one of the most popular. LSTM is here to complement the shortcomings of RNN. It cannot predict words based on historical information stored for a long period of time. [17] Therefore, LSTM can remember long-term stored aggregated information and remove irrelevant information. LSTM is more efficient when processing, predicting, and classifying data based on specific time series. The gate studies weights that control the decay rate of values stored in memory cells. [18] For example, if the input and output gates are off and there is no rollover caused by the forget gate, the memory cell will retain its value over time, so the slope of the error will remain constant as long as the backward propagation increases. This allows the model to remember information longer. Mathematically, each step can be described as: In the first step, the forget gate layer sees 1 and new inputs to decide which features to remove from the cell state.
In the second step, the decision on what information is stored in the state of the cell is made in two steps. The output gate layer, which is a sigmoid layer, specifies the value to be updated. Then the light brown layer creates a new candidate value vector.
The old cell state Ct−1 is updated to the new cell Ct summing the output of the forget gate layer functions ft and it × Ct.
The output is specified in two steps -First, the sigmoid layer decides the parts of the cell to be ejected. The product of the new cell state Ct through tan h and the output from the sigmoid gate produces a selectively defined h t part.
Tuning and optimization of hyperparameters is a difficult and experimental task. LSTM model training is expensive in terms of memory and computing power.

Stages of Research
From the flowchart in Figure 2.2 under. The first step is Data Collection. The data is a collection of PQM variables in PT. Semen Baturaja (Persero) Tbk. Then these variables are processed so that which variables are selected which influence the prediction of power will later be carried out. Then the dataset is divided into two, namely Data Train and Data Test. Data Train will train the dataset using the LSTM algorithm and the Data Test plays a role in testing the training results that have been carried out by the Data Train so that the output results can be in the form of a test score, train score, graph, and loss value per epoch.

Results and Discussion
To make predictions in real time, you need pseudocode as stated in table 3.1 below: Preparing the scaler Step-2 Fixing random seeds to increase productivity Step-3 Enter the monitoring dataset and enter the company website URL Step-4 Enter the variable you want to predict Step-5 Entering the LSTM algorithm and making predictions Step- 6 Print prediction results Before entering the dataset in the algorithm that has been prepared, it's a good idea to prepare the scaler first, scaling aims to make the numerical data on the dataset have the same range of values (scale). There is no longer one data variable dominating the other data variables and then proceeds to fix random seeds to increase predictive productivity. [19][20]    The accuracy of online predictions on active power A reached 87.72%, which is certainly a good result because it exceeds 50% of the expected results.

Realtime VS Actual Active Power B Prediction Results
The value of the difference in power prediction on active power A between 13.00 -14.  The accuracy of online predictions on active power B reached 89.27%, which is even better than the previous predictions of active power A.

Realtime VS Actual Active Power C Prediction Results
The value of the difference in power prediction on active power A between 13.00 -14.00 is 3770. 16

Realtime VS Actual Active Power Prediction Results
The value of the difference in power prediction on active power A between 13.