Research: Prediction of Energy Consumption for Machinery in Manufacturing Process
Energy plays a vital role in the manufacturing process in packaging films industries to produce high-quality products. Consumption of energy plays a key role in the cost of polyester film products. In manufacturing industries, optimal energy leads to a higher production volume of products. Currently, the plant is facing high energy consumption at various times which is leading to increased power consumption and turn into higher cost of production. It leads to frequent breakdown of maintenance activities which causes an impact on productivity. This paper aims to analyze and identity machines that are consuming abnormal power in the manufacturing process in the given period. Also, this paper will perform future production of power consumption for a machine.
I. INTRODUCTION
Power Energy plays a key role in manufacturing process in package films industries to produce with high quality product. The early research, headed by W.H. Carothers, centered on what became nylon, the primary fiber [1]. Nowadays, there are two predominant types of polyester, PET (polyethylene terephthalate) [2] and PCT (poly-1, 4-cyclohexylene-dimethylene terephthalate). Biaxially-oriented polyethylene terephthalate (BoPET)[2] is a polyester film made from polyethylene terephthalate. The film is heat-set under tension in the stretching oven at temperatures which are typically above 200 °C. BOPP films (Biaxially Oriented Polypropylene Films) [3] are produced by stretching polypropylene film in both machine direction and transverse direction. BOPP film is used in many applications comprising packaging, labeling and lamination. BoPET Film manufacturing process involved below steps:
- Feeding
- Crystallization
- Drying
- Extrusion and Die
- Casting
- Machine Direction Orienter
- Transverse Direction Orienter
- Take up and transfer
- Winder
In this paper, we will be predicting the power consumption of the machinery tools which are used in polyester film manufacturing industry based on historical data which has been collected from the respective machinery over the period of time.
Since the 1950s, many previous studies can be found in the literature on melting and thermal issues in polymer extrusion till 1990. Post 1990, researchers are focused on power energy consumption in polymer manufacturing process. Many industries are spending more cost for power energy in few decades. Hassan, Mohamed Ket al (2013) researched [4] on real-time stress-strain birefringence measurements to interpret the sequence of mechanism that occur during biaxial stretching of PET films. The measurement of above research results in four ordered way that determines the relationship between stress and birefringence. The First 2 results have a positive linear relationship between stress and the ‘Rays of lights’(birefringence) whereas third resulted in a non-linear relationship. The fourth stage, the stress is highly increased while the birefringence reaches the plateau and this stage is called finite extensibilities. The initial linear stress test coincides with the onset of the stress induces crystallization by the offline measurements.
Nisticò, Roberto (2020) has published an article [5] on Polymer Testing using Polyethylene terephthalate (PET). The PET is 3rd most widely used polymer in the packaging industry. The PET is primarily derived from fossil sources and it is nonbiodegradable in the environment. The research identifies there is a possibility of manufacture the PET material using biodegradable material in a more sustainable way. This allows the environment safe and re-use /recycle of PET material and makes indefinitely use of PET material in the manufacturing industries. Author has discussed more on PET usage in the food packaging and analyze advantage and disadvantage of the PET material.
In manufacturing process, temperature is key factor for producing high quality of product. Nowadays, it is challenging to maintain both melt and energy efficiency in same time. There are three type of screws BF, GC and RC in the machine and speed of these screws helping to reduce power consumption of machine. If speed of screw increases, thermal fluctuation also increases [6] [7].
In all lines, the company continues to optimize constantly the possibilities for energy savings. The development of the fast-running high-speed extrusion (HSE) single screw extruder, especially as co-extruder for film stretching lines, is aimed to achieve the performance of several larger conventional extruders with considerably smaller screw diameters. The HSE extruder is easy to operate and less expensive to service and requires smaller installation area. With regards to energy consumption a small extruder also provides advantages because it has lower radiation and convection losses. The direct drive contributes to the lower energy consumption since there are no energy losses resulting from the reduction gear. With regards to the process technology, the new extruder concept also leads to shorter dwell times of the melt in the extruder. This enables improvement of the film characteristics, purging and changeover times can be reduced.
According to research paper done by Girish Kant [8], build Artificial Neural Network(ANN) and Support Vector Regression(SVR) models to predict the model for power consumption of the machine which resulted that there is close correlation between predicted power and actual experimental power consumption. ANN is performed well to predict power consumption when compare to SVR model prediction. The predictive models are expected to help in arriving at fine-tuned optimum machine parameters so that power consumption during machining can be reduced.
Mel Keytingan M.Shapi published the paper [9] which stats that power consumption can be predicted using K-Nearest Neighbor, Azure ML, Support Vector Machine (SVM), ANN which resulted 90% accuracy and lower rate of bias. Since they recommended for further analysis with more powerful system or platform to run the SVM algorithm. Also, more data variables to increase the performance of the model.
II. MATERIALS AND METHODS
A. Data Acquisition and Pre-processing
The dataset has been collected for below listed machine at various stages in film manufacturing production machine unit.
- Chilled Water
- Cooling Water Pump
- Cooling Tower Fan
- Air Compressor
- AHU 6 & 7
- HVAC Panel Room No.1
- HVAC Panel Room No.2
- HAVAC Panel Room No.3&4
TABLE 1: LIST OF ATTRIBUTES FOR WHICH DATA COLLECTED FOR THE ABOVE MACHINES
| S. No | Attributes | Description |
|---|---|---|
| 1 | Date | The machine operated date |
| 2 | Dry Bulb | Ambient Temperature Outside |
| 3 | Wet bulb | How much precipitation in the atmosphere |
| 4 | RH% | Relative humidity calculated based on Dry bulb and web Bulb |
| 5 | Number of pumps running | Number of pumps used |
| 6 | Screw Chiller 1,2,3 & 4 | There are four screw chiller machine and each machine having two compressors |
| 7 | Running Hours | How long each compressor is running in a day. |
| 8 | Power Consumption | How much power each screw chiller type machine consumed per day. |
| 9 | Header Pressure | Pressure of Chilled water at the outlet |
| 10 | Supply Temperate | Temp of the Chilled water at the outlet |
| 11 | Return Temperate | Temp of chilled water at the Inlet of the Chiller |
| 12 | Supply header flow | Flow of the Chilled water |
| 13 | Chiller load TR Only Load % | Range in which Chiller if working |
| 14 | Booster pump discharge pressure | Pressure of chilled water going through machines |
| 15 | Total Energy Consumption | It’s an overall energy consumed by all screw chiller machine 1, 2, 3 & 4 |
Feature attributes for this prediction will use electrical power data consisting of Date, Dry/Wet bulb temperature, Header pressure to inlet and outlet and power consumption, in which the power consumption would be the targeted output.
Most of the time, raw data is not complete and it cannot be sent for processing (applying models). Here, preprocessing the dataset makes it suitable to apply analysis on. This is an extremely important phase as the final results are completely dependent on the quality of the data supplied to the model. However great the implementation or design of the model is, the dataset is going to be the distinguishing factor between obtaining excellent results or not.
- Pre-processing steps like impute missing values, handling outliers, one-hot encoding the categorical variables and scale features using standard scalar.
- Impute Missing values with zeros in the independent feature.
- Check out for time exceeding max limit of 24 hours /day in the independent features for each machine running hours.
The training and test data are preprocessed with the above methods. Once the preprocessing steps applied, the data is ready to be used in model.
The imputation method was evaluated to determine its performance. The resultant data which has been cleaned then further pre-processed using Z-score normalization. It is a transformation to change the observed data to have characteristics of standard normal distribution in which the mean is 0, and the standard deviation is 1. This transformed the data to be equally distributed above and below the mean value by using the formula in equation.
Where z is scaled value and x is a feature.
B. Feature Extraction
After loading the data into data frame, an extensive feature extraction method was applied, and this features extraction technique narrowed down the features necessary for building regression model. As it is time series, we concreate the independent and dependent variables with window size 3. Last 10 days data considered as test set and remaining data considered as training set.
C. Model
This research paper uses time series algorithm to predict the power consumption. After data is pre-processed, it is then applied to build a predictive model. Before using the data to train the model, data partitioning was done to separate the data into two groups – a training group and a testing group.
The research was performed using 3 types of time series ML algorithm [10] – Long short-term memory (LSTM), Autoregressive integrated moving average (ARIMA) and Facebook Prophet.
LSTM algorithm [11][20] is used in time series predictions due to its sequential prediction methodology. The LSTM helps in mitigate exploding and vanishing gradient. The LSTM algorithm gives a better prediction compared to other basic algorithm like linear regression.
Each time step of the test dataset will be walked one at a time. A model will be used to make a forecast for the time step, then the actual expected value from the test set will be taken and made available to the model for the forecast on the next time step.
The prediction on the test dataset will be collected and an error score (RMSE and MAPE) calculated to summarize the outcome of the model.
For this research, LSTM algorithm is used by considering “DATE” as independent feature and “Power Consumption” as dependent feature. From the dataset, last 10 days data is used as test set and remaining data as training set.
LSTM algorithm was used to predict power consumption for all 8 machines and the resultant model from the testing with the lowest MAPE value was chosen for final results.
ARIMA algorithm[13][14] is widely used for predicting future trends on time series data. Before applying ARIMA algorithm, the data should be verified for non-stationary and seasonal component. In this research, the data for last 5 months has non-stationary and seasonality in it. We use an ARIMA(p,d,q) model for time series data yt, where p is the number of autoregressive lags[12], d is the degree of differencing and q is the number of moving average lags as shown in equation (1).
The appropriate lag is obtained through autocorrelation plot and same has been used in ARIMA model to the latent variable estimates and forecast power consumption of a polyester film machine for next 10 days. The ARIMA model summary and residuals are captured and the latent variables are analyzed.
The third methodology used in this research for power consumption is Facebook Prophet [15] [19]. The advantage of using Prophet is that it works best with time series that have strong seasonal effects and several seasons of historical data. Prophet is robust in handling missing data, shift in the trend and handle outliers well. Works well with time-series data having non-linearity in trends as well as holiday effects.
The input to Prophet is always a dataframe with two columns: ds and y. The “DATE” feature is defined as ds in a format YYYY-MM-DD and the “Power Consumption” feature is considered as y which must be numeric and represents the measurement we wish to forecast.
The hyper parameters can be tuned with the help of “ParameterGrid” function to find best parameter value to find the prediction accurately. The hyperparameter tuned for this research are ‘number of changepoint’, ‘seasonality mode’ and ‘changepoint prior scale’. We can get a suitable dataframe that extends into the future a specified number of days using the helper method ”make_future_dataframe”.
In this research we have used different methodology and used different set of training and testing data. For the LSTM model, the data was partitioned into two groups whereby 80% of the dataset was used for training and the other 20% was partitioned as testing data groups. For the ARIMA and Prophet model, the last 10 days data was used as test set and remaining data as training data.
The training groups of data were used to train each machine learning algorithm and generate a predictive model that could output value that matches with the recorded maximum data while the remaining data was held back to be used to test the trained predictive model.
During model training, various models were created with different tuning parameters. In which ‘learning rate’ and ‘optimizer’ was adjusted for LSTM model; parameters ‘p’,’d’,’q’ values are adjusted for ARIMA model[16] and ‘number of changepoint’, ‘seasonality mode’, ‘changepoint prior scale’ parameters are adjusted for Prophet model.
After the repeated tuning finished up to its respective maximum parameters, each model was evaluated based on Root Mean Square Error (RMSE) [17], R-Squared (R2) and Mean Average Error (MAE) [17]. The formula is as shown in equations, given that At is the actual recorded values of maximum demand data, and Ft is the predicted values. Although 3 evaluations were made, only RMSE and MAPE [18] result was acknowledged as the best model for each method.
After the predictive model using each machine learning algorithm was developed and prediction data was generated, they were then evaluated to determine their performance and accuracy. Two methods of evaluation were used which were Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). The formula for RMSE and MAPE are shown in equation.
III. RESULTS
The results of the experimentation were discussed in this section based on the steps of the predictions’ framework. The findings regarding power consumption prediction were reviewed for each machine and performance comparison was provided for the prediction result of LSTM, ARIMA, Prophet and Linear regression.
The MAPE metrics was used to evaluate the model performance and accuracy. Power consumption for each machine derives a different result based on algorithm used. Facebook Prophet and ARIMA algorithm derives best accuracy and prediction of power consumption.
TABLE 2: MODEL MAPE VALUE
| Machine Name | Model | MAPE |
|---|---|---|
| Chilled Water | Facebook Prophet | 6.02 |
| ARIMA | 7.55 | |
| LSTM | 7.69 | |
| Linear Regression | 8.57 | |
| Cooling Water Pump | Facebook Prophet | 2.87 |
| LSTM | 3.18 | |
| ARIMA | 4.43 | |
| Air Compressor | Facebook Prophet | 1.94 |
| ARIMA | 3.47 | |
| LSTM | 3.73 | |
| Cooling Tower | ARIMA | 10.33 |
| Facebook Prophet | 11.15 | |
| LSTM | 14.15 | |
| HVAC-1 | ARIMA | 2.27 |
| Facebook Prophet | 2.54 | |
| LSTM | 3.58 | |
| HVAC-2 | Facebook Prophet | 3.24 |
| LSTM | 3.62 | |
| ARIMA | 4.27 | |
| HVAC-3 & 4 | ARIMA | 3.64 |
| Lagged linear | 7.95 | |
| LSTM | 22.45 | |
| AHU | Facebook Prophet | 1.77 |
| ARIMA | 2.29 | |
| LSTM | 2.92 |
Pictorial representation of power consumption prediction for each machine in the polyester film manufacturing process.
IV. DISCUSSION AND CONCLUSION
This research has focused on developing a power consumption predictive model for eight machine types that are being used in polyester film manufacturing industry. The power consumption data collected from the polyester film manufacturing plant has been analyzed and pre-processed for training and testing of the predictive models.
A statistical analysis of the data collected was made to determine the power consumption per hour. From this analysis, a new feature was identified as “Total Running hours” for each machine. As part of predictive model development, imputation of missing data was carried out in the data collected from machines in the manufacturing plants.
In this research, three time series machine learning prediction regression methods namely LSTM, ARIMA and Facebook Prophet were chosen as the algorithm for the predictive model. These methods were successfully compared in terms of their prediction performances. The consequence of the model training and testing shows that each method performed differently for every machine types. In this research Facebook Prophet algorithm provided better accuracy when compared to other time series algorithms.
There are a few improvements that can be done to improve this research. There was a challenge in collecting real-time data with sufficient attributes from polyester film manufacturing industry. The model accuracy would be derived better if good amount of real-time data with sufficient attributes is provided.
V. REFERENCES
- W. H. Carothers and F. J. van Natta, “Studies of polymerization and ring formation. XVIII. Polyesters from ι-hydroxydecanoic acid,” J. Am. Chem. Soc., vol. 55, no. 11, pp. 4714–4719, 1933.
- J. Breil, “Future trends for biaxially oriented films and orienting lines,” in Biaxial Stretching of Film, Elsevier, 2011, pp. 240–273
- W. Xia, J. Lu, S. Tan, J. Liu, and Z. Zhang, “Manipulating dielectric properties by modifying molecular structure of polymers,” in Dielectric Polymer M
- M. K. Hassan and M. Cakmak, “Mechano optical behavior of polyethylene terephthalate films during simultaneous biaxial stretching: Real time measurements with an instrumented system,” Polymer (Guildf.), vol. 54, no. 23, pp. 6463–6470, 2013.
- Nisticò, R. (2020). Polyethylene terephthalate (PET) in the packaging industry. Polymer Testing, 90(106707), 106707. https://doi.org/10.1016/j.polymertesting.2020.106707
- Abeykoon, C., Kelly, A. L., Vera-Sorroche, J., Brown, E. C., Coates, P. D., Deng, J., Li, K., Harkin-Jones, E., & Price, M. (2014). Process efficiency in polymer extrusion: Correlation between the energy demand and melt thermal stability. Applied Energy, 135, 560–571. https://doi.org/10.1016/j.apenergy.2014.08.086
- Abeykoon, C., Kelly, A. L., Brown, E. C., Vera-Sorroche, J., Coates, P. D., Harkin-Jones, E., Howell, K. B., Deng, J., Li, K., & Price, M. (2014). Investigation of the process energy demand in polymer extrusion: A brief review and an experimental study. Applied Energy, 136, 726–737. https://doi.org/10.1016/j.apenergy.2014.09.024
- Girish Kanta, Kuldip Singh Sangwan, (2015) Predictive Modelling for Power Consumption in Machining using Artificial Intelligence Techniques. Procedia CIRP 26 (2015) 403 – 407
- Shapi, M. K. M., Ramli, N. A., & Awalin, L. J. (2021). Energy consumption prediction by using machine learning for smart building: Case study in Malaysia. Developments in the Built Environment, 5(100037), 100037. https://doi.org/10.1016/j.dibe.2020.100037
- Verma, P., Reddy, S. V., Ragha, L., & Datta, D. (2021). Comparison of time-series forecasting models. 2021 International Conference on Intelligent Technologies (CONIT), 1–7.
- G. Li, X. Zhao, C. Fan, X. Fang, F. Li, and Y. Wu, “Assessment of long short-term memory and its modifications for enhanced short-term building energy predictions,” J. Build. Eng., vol. 43, no. 103182, p. 103182, 2021.
- N. Darapaneni et al., “Food temperature analysis and forecasting,” in 2021 IEEE World AI IoT Congress (AIIoT), 2021, pp. 0143–0149.
- B.-M. Hodge, A. Zeiler, D. Brooks, G. Blau, J. Pekny, and G. Reklatis, “Improved wind power forecasting with ARIMA models,” in Computer Aided Chemical Engineering, vol. 29, E. N. Pistikopoulos, M. C. Georgiadis, and A. C. Kokossis, Eds. Elsevier, 2011, pp. 1789–1793.
- Elsaraiti, M., Ali, G., Musbah, H., Merabet, A., & Little, T. (2021). Time series analysis of electricity consumption forecasting using ARIMA model. 2021 IEEE Green Technologies Conference (GreenTech), 259–262.
- R. J. Chadalavada, Department of CSE, School of Engineering and Technology, CHRIST Deemed to be University, Mysore Road, Bengaluru, 560074, India, S. Raghavendra, V. Rekha, Department of CSE, School of Engineering and Technology, CHRIST Deemed to be University, Mysore Road, Bengaluru, 560074, India., and Department of CSE, School of Engineering and Technology, CHRIST Deemed to be University, Mysore Road, Bengaluru, 560074, India, “Electricity requirement prediction using time series and Facebook” DOI: 10.17485/IJST/v13i47.1847 Year: 2020, Volume: 13, Issue: 47, Pages: 4631-4645
- N. Darapaneni, D. Reddy, A. R. Paduri, P. Acharya, and H. S. Nithin, “Forecasting of COVID-19 in India using ARIMA model,” in 2020 11th IEEE Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON), 2020, pp. 0894–0899.
- T. Chai and R. R. Draxler, ``Root mean square error (RMSE) or mean absolute error (MAE)? _Arguments against avoiding RMSE in the literature,''Geosci. Model Develop., vol. 7, pp. 1247_1250, Jun. 2014. doi: 10.5194/gmd-7-1247-2014.
- S. Kim and H. Kim, “A new metric of absolute percentage error for intermittent demand forecasts,” Int. J. Forecast., vol. 32, no. 3, pp. 669–679, 2016.
- C. B. Aditya Satrio, W. Darmawan, B. U. Nadia, and N. Hanafiah, “Time series analysis and forecasting of coronavirus disease in Indonesia using ARIMA model and PROPHET,” Procedia Comput.
- Asri, M. A. N. M., Zaini, N., & Latip, M. F. A. (2021). Development of an LSTM-based model for energy consumption prediction with data pre-analysis. 2021 11th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 228–233.