PaperlayerView exact revision on arXiv

Source-linked AI summary

A Data Mining Approach Combining K-Means Clustering with Bagging Neural Network for Short-term Wind Power Forecasting

Wenbin Wu, Mugen Peng

arXiv:1709.05261v1cs.CE

TL;DR

Wind power’s intermittency and volatility make forecasting accuracy important, while diverse training samples complicate short-term forecasting. The paper clusters historical days by meteorological conditions and power, then uses bagging with a BPNN. On real wind-generation data, it reports better accuracy than baseline and other short-term forecasting approaches.

Abstract

from arXiv · show

Wind power forecasting (WPF) is significant to guide the dispatching of grid and the production planning of wind farm effectively. The intermittency and volatility of wind leading to the diversity of the training samples have a major impact on the forecasting accuracy. In this paper, to deal with the training samples dynamics and improve the forecasting accuracy, a data mining approach consisting of K-means clustering and bagging neural network is proposed for short-term WPF. Based on the similarity among historical days, K-means clustering is used to classify the samples into several categories, which contain the information of meteorological conditions and historical power data. In order to overcome the over fitting and instability problems of conventional networks, a bagging-based ensemble approach is integrated into the back propagation neural network. To confirm the effectiveness, the proposed data mining approach is examined on real wind generation data traces. The simulation results show that it can obtain better forecasting accuracy than other baseline and existed short-term WPF approaches.

I. INTRODUCTION

Wind power’s intermittency makes accurate forecasting important for grid security and reliability. The paper combines K-means clustering with bagging neural networks, addressing training-sample selection as well as network instability and overfitting.

I. INTRODUCTION

  • Wind power’s fluctuating and intermittent output makes accurate forecasting important for grid security, reliability management, and spinning-reserve costs.Forecasting supports grid dispatch and wind-farm production planning.Source: §I · p. 1
  • Existing WPF approaches include physical, statistical, and combination methods, with physical-model accuracy depending on available site information.Physical approaches also require detailed turbine and wind-farm characteristics and have limited versatility across sites and operating states.Source: §I · p. 1
  • Prior BPNN improvements address parameters, architecture, or kernels, while earlier clustering approaches select training data without considering turbine operating state.The paper identifies training-sample selection as an additional factor affecting forecast accuracy.Source: §I · pp. 1–2
  • The paper proposes K-means clustering and a bagging neural network to forecast short-term wind power for an individual turbine.The contributions describe clustering by meteorological conditions and historical power, then averaging outputs from bootstrapped BPNNs.Source: §I · p. 2
  • The proposed approach reports lower RMSE and MAE than baseline approaches, indicating improved forecasting accuracy.It combines K-means clustering with a bagging-based BPNN.Source: §I · p. 2

II. THE ESTABLISHMENT OF THE WIND POWER FORECASTING MODEL

The forecasting model preprocesses the data, selects training samples similar to the forecast day through K-means clustering, and predicts wind power with a bagging neural network.

II. THE ESTABLISHMENT OF THE WIND POWER FORECASTING MODEL

  • Data preprocessing cleans and normalizes samples and selects variables related to wind power as neural-network inputs.The model description presents these operations as the first stage of the pipeline.Source: §II · p. 2
  • K-means clustering selects the training set most similar to the forecasting day from the preprocessed data.Source: §II · p. 2
  • A bagging neural network forecasts wind power and is intended to alleviate BPNN instability and overfitting.Source: §II · p. 2

A. Data preprocessing

Preprocessing removes unreasonable observations, selects relevant features, and normalizes values before forecasting.

A. Data preprocessing

  • Preprocessing is used to address unreasonable samples, irrelevant or redundant features, and normalization’s effects on convergence and accuracy.Source: A. Data preprocessing · p. 2
  • Negative wind-speed or wind-power values are removed, and deleted values are filled using the mean of the preceding and following data.Source: A. Data preprocessing · p. 2
  • Relief weights features by correlation and removes features below a threshold; its running time grows linearly with sampling times and original features.The passage also states that Relief reduces physical dimensions compared with PCA.Source: A. Data preprocessing · pp. 2–3
  • Min-max normalization maps data to [0,1] using the training-set minimum and maximum, with the stated aim of improving convergence rate and forecasting accuracy.Source: A. Data preprocessing · p. 3

B. Similar day clustering based on meteorological conditions and historical power

The clustering method selects training days using similarity in both meteorological conditions and historical power, then applies K-means to form groups of similar samples.

B. Similar day clustering based on meteorological conditions and historical power

  • Meteorological samples use wind-speed and temperature summaries, while historical-power summaries are also included because similar weather may not imply similar power.The passage attributes this difference to geographical conditions and turbine operating state.Source: B. Similar day clustering based on… · p. 3
  • Because the forecast day’s power summary is unknown, the method assumes the preceding day has the same summary and clusters meteorological and power samples separately.Days from the clusters most similar to the forecast day are combined as the training samples.Source: B. Similar day clustering based on… · p. 3
  • K-means assigns data to the nearest randomly selected center and updates each center to its cluster’s mean until it stops changing.Source: B. Similar day clustering based on… · p. 3
  • Pearson correlation is used to identify the class most related to the forecasting day for neural-network input.Source: B. Similar day clustering based on… · p. 4

C. Bagging neural network

The paper uses bagging to train multiple BPNNs on bootstrap samples and average their forecasts equally. This ensemble is intended to address BPNN instability and overfitting while improving forecasting accuracy and reducing time costs.

C. Bagging neural network

  • BPNN propagates data forward and then backpropagates error, using gradient descent to update weights and minimize the error function.The repeated weight modifications consume time.Source: C. Bagging neural network · p. 4
  • The network is tested with M from 2 through 10 input neurons to find the best forecasting result.Source: C. Bagging neural network · p. 4
  • BPNN can trap into local optimization, become unstable when training samples change little, and overfit; bagging is introduced to address these problems.Source: C. Bagging neural network · p. 4
  • Bagging trains BPNNs on K randomly sampled subsets and averages their forecasts, with all models assigned equal importance.The process begins with bootstrap sampling of the training data before applying the learning algorithm to subsets.Source: C. Bagging neural network · pp. 4–5
  • Bagging is integrated into BPNN forecasting to improve accuracy and reduce time costs.The paper also motivates ensemble learning as a way to enhance model stability and forecasting accuracy.Source: C. Bagging neural network · pp. 4–5

III. SIMULATION RESULTS

The simulation uses wind-farm SCADA data and evaluates forecasts with RMSE and MAE. The 10-minute readings are aggregated into hourly data by summing six readings.

III. SIMULATION RESULTS

  • The study uses 10-minute data from a wind farm’s SCADA system, summing six readings to form hourly data.Source: §III · p. 5
  • Forecast performance is evaluated with RMSE and MAE, using actual power, predictive power, and the number of test data points.Source: §III · p. 5

A. Data preparing

The data preparation selects forecast inputs and groups historical days by meteorological conditions and power data. The most related category supplies the training days for forecasting.

A. Data preparing

  • Power generation curves differed sharply before and after July and varied even within seasons, motivating clustering.Fig. 4 presents the wind power generation curve from December 2013 to November 2014.Source: A. Data preparing · p. 5, Figure 4 · p. 5
  • Average wind speed, blade angle, and ambient temperature are selected as inputs; blade angle represents wind direction and is encoded with sine and cosine values.Relief ranks wind speed highly, while machine and generator temperatures are unavailable on the forecast day.Source: A. Data preparing · p. 5
  • Each day is a 144-item training unit, clustered using daily wind-speed and temperature summaries plus power-generation totals and extrema.Source: A. Data preparing · p. 6
  • K-means separates the initial samples into three groups with distinct wind-speed, temperature, and power profiles; Pearson correlation then identifies the most related category for training.The selected category contributes 29 training days.Source: A. Data preparing · p. 6
  • Table II reports the performance of baseline approaches and the proposed approach.Source: Table II · p. 6

B. Forecasting results

The proposed approach is compared with BPNN, BPNN with clustering, and other forecasting methods. It reports lower forecasting errors than these comparison methods.

B. Forecasting results

  • The proposal reduced RMSE by 38.7% and 29.8% compared with BPNN, and MAE by 30.1% and 21.1% compared with BPNN with clustering.The prediction results are reported in Table II.Source: B. Forecasting results · p. 6
  • With clustering, BPNN reduced RMSE by 12.7% and MAE by 11.5% compared with BPNN without clustering.The proposal was reported as the best of the three approaches.Source: B. Forecasting results · p. 6
  • The proposal had the best forecasting accuracy among the Approach 1, Approach 2, and Approach 3 baselines; it reduced RMSE and MAE by 19.9% and 17.1% versus Approach 2.Approach 3 outperformed Approaches 1 and 2 but remained below the proposal.Source: B. Forecasting results · pp. 6–7, Table III · p. 7

IV. CONCLUSION

The proposed K-means and bagging-neural-network approach uses meteorological conditions and historical power to cluster data and improve short-term wind-power forecasting. The authors report significant RMSE and MAE gains, while identifying farm-scale application and further model development as important directions.

IV. CONCLUSION

  • The proposed approach combines K-means clustering and a bagging neural network, using meteorological conditions and historical power to group data and improve forecast stability and accuracy.Pearson correlation measures the forecasting day's distance from clusters, and Relief selects neural-network inputs.Source: §IV · p. 7
  • The authors report significant gains in RMSE and MAE on actual data from a practical wind farm.The passage does not provide metric values or identify specific comparison baselines.Source: §IV · p. 7
  • The approach is not designed for individual turbines, but the authors propose that clustering could extend to large wind farms and reduce computational complexity while increasing forecast accuracy.They describe grouping turbine wind speeds when turbine locations lie in one direction.Source: §IV · p. 7
  • Future work should address effective meteorological forecasting, BPNN optimization, multidimensional clustering, and wind-farm forecasting models.Source: §IV · p. 7
Loading 1709.05261v1…