Scholarlink

<p>Univariate time series (TS) method forecasts the future series of a single dependent variable using time as an independent variable. Hence, it aids in analyzing the underlying dynamics and forecasting the trend of humidity as well as predicting the effects of humidity on different facets of our lives. TS data can be modelled by statistical and machine learning (ML) methods. However, statistical methods are productive for modeling linear components of TS data whereas ML methods are effective for modelling its non-linear components. Further, the prediction accuracy of TS models deteriorates due to presence of anomalies in data. To address such limitations, we have proposed a hybrid univariate TS model for prediction of TS relative humidity (RH) data by merging K-Nearest Neighbour (KNN), statistical Multiplicative Holt-Winter Seasonal (MHWS), and ML based neural network autoregression (NNAR) method. In the proposed model, anomaly free RH data produced by KNN model are input to MHWS model for forecasting RH series and its generated residuals are input to NNAR model for forecasting error series. The predicted RH series and error series by MHWS and NNAR models respectively are combined to produce future RH timeseries data. To prove the effectiveness of KNN, MHWS, and NNAR models, performance comparison is made for (i) univariate point anomaly detection methods: K-Means, KNN, Gaussian Mixture Model (GMM) and One-Class Support Vector Machine (OC-SVM); (ii) statistical methods: MHWS, additive HWS, and Autoregressive Integrated Moving Average (ARIMA); (iii) ML methods: NNAR and Random Forest (RF). The proposed model is robust in prediction of RH timeseries with estimated RMSE = 3.1. The proposed hybrid model can be used for the applications where individual data series are focused primarily.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid KNN-MHWS-NNAR Univariate time Series Model for Anomaly Detection and Prediction of Relative Humidity Data

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Bishnupriya Mallick,
  • Jibendu Kumar Mantri

摘要

Univariate time series (TS) method forecasts the future series of a single dependent variable using time as an independent variable. Hence, it aids in analyzing the underlying dynamics and forecasting the trend of humidity as well as predicting the effects of humidity on different facets of our lives. TS data can be modelled by statistical and machine learning (ML) methods. However, statistical methods are productive for modeling linear components of TS data whereas ML methods are effective for modelling its non-linear components. Further, the prediction accuracy of TS models deteriorates due to presence of anomalies in data. To address such limitations, we have proposed a hybrid univariate TS model for prediction of TS relative humidity (RH) data by merging K-Nearest Neighbour (KNN), statistical Multiplicative Holt-Winter Seasonal (MHWS), and ML based neural network autoregression (NNAR) method. In the proposed model, anomaly free RH data produced by KNN model are input to MHWS model for forecasting RH series and its generated residuals are input to NNAR model for forecasting error series. The predicted RH series and error series by MHWS and NNAR models respectively are combined to produce future RH timeseries data. To prove the effectiveness of KNN, MHWS, and NNAR models, performance comparison is made for (i) univariate point anomaly detection methods: K-Means, KNN, Gaussian Mixture Model (GMM) and One-Class Support Vector Machine (OC-SVM); (ii) statistical methods: MHWS, additive HWS, and Autoregressive Integrated Moving Average (ARIMA); (iii) ML methods: NNAR and Random Forest (RF). The proposed model is robust in prediction of RH timeseries with estimated RMSE = 3.1. The proposed hybrid model can be used for the applications where individual data series are focused primarily.