💥1 概述
本研究的目的是预测压缩空气系统的电气负载曲线，这对于行业从业者和软件提供商开发更好的负载管理和前瞻调度程序的实践和工具很有价值。使用两层前馈神经网络和长短期记忆两个人工神经网络来预测空压机的电气负荷。
📚2 运行结果
2.1 算例1
2.2 算例2
2.3 算例3
部分代码：
close all clear all % load data load('data/ac1Data.mat') % assign data to variables Load = ac1.load; Weekday = weekday(ac1.DATE); Minute = minute(ac1.Time); Hour = hour(ac1.Time); Pressure = ac1.LinePressure_bar_; Temp1 = ac1.InletAirTempC; Temp2 = ac1.DischargeTemperature_degC_; % form input exoData = [Weekday,Minute,... Hour,Pressure,Temp1,Temp2]; % define working day index workingDayIdx = and(Weekday >1, Weekday < 7); % define non-working day index nonworkingDayIdx = or(Weekday ==1,Weekday == 7); % data pre-process/apply hampel filter workingExoData = exoData(workingDayIdx,:); workingLoad= hampel(Load(workingDayIdx),5,2); nonworkingExoData = exoData(nonworkingDayIdx,:); nonworkingLoad = hampel(Load(nonworkingDayIdx),5,2); % define working day test set inputTestSeries1 = num2cell(workingExoData(end-200:end,:)',1); targetTestSeries1 = num2cell(workingLoad(end-200:end)'); % define non-working day test set inputTestSeries2 = num2cell(nonworkingExoData(end-200:end,:)',1); targetTestSeries2 = num2cell(nonworkingLoad(end-200:end)'); % load pre-trained networks load('data/ac1_w_ffnn.mat') % for working days load('data/ac1_nw_ffnn.mat') % for non-working days netsw = removedelay(netw); netsn = removedelay(netn); % forecast values using pre-trained networks [xsw,xisw,aisw,tsw] = preparets(netsw,inputTestSeries1,{},targetTestSeries1); ys_w = netsw(xsw,xisw,aisw); [xsn,xisn,aisn,tsn] = preparets(netsn,inputTestSeries2,{},targetTestSeries2); ys_n = netsn(xsn,xisn,aisn); % timeshfit forecasts back to the original positive -- see 'removedelay' documentation yhat_w = cell2mat(ys_w(end-96:end-1)); y_w = workingLoad(end-(length(yhat_w)-1):end)'; yhat_n = cell2mat(ys_n(end-96:end-1)); y_n = nonworkingLoad(end-(length(yhat_n)-1):end)'; % calculate evaluation metrics workingDayRMSE = sqrt(mean((y_w - yhat_w).^2)); % calculate root mean squared error for working day forcast nonworkingDayRMSE = sqrt(mean((y_n - yhat_n).^2)); % calculate root mean squared error for non-working day forcast RMSE = [workingDayRMSE;nonworkingDayRMSE]; workingDayMASE = mean(abs(y_w-yhat_w))/(mean(abs(y_w(2:end)-y_w(1:end-1)))); % calculate mean absolute scaled error for working day forecast nonworkingDayMASE = mean(abs(y_n-yhat_n))/(mean(abs(y_n(2:end)-y_n(1:end-1)))); % calculate mean absolute scaled error for non-working day forecast MASE = [workingDayMASE;nonworkingDayMASE]; mdl_w = fitlm(y_w,yhat_w); workingDayR2 = mdl_w.Rsquared.Ordinary; % get R2 between observed and predicited for working day forecast mdl_n = fitlm(y_n,yhat_n); nonWorkingDayR2 = mdl_n.Rsquared.Ordinary ;% get R2 between observed and predicited for non-working day forecast R2 = [workingDayR2;nonWorkingDayR2]; T = table (RMSE,MASE,R2,'RowNames',{'Working Days';'Non-working Days'}); % construct output table T.Properties.DimensionNames{1} = 'Mode'; figure subplot(2,2,1) plot(y_w) hold on plot(yhat_w,'.-') hold off legend(["Measured" "Predicted"]) xlabel("Timestep (15-minutes)") ylabel("Electrical Load (kW)") title(["Forecast using FFNN";"Compressor 1 - Working Day"]) subplot(2,2,3) stem(yhat_w - y_w) xlabel("Timestep (15-minutes)") ylabel("Error (kW)") title("RMSE = " + workingDayRMSE) subplot(2,2,2) plot(y_n) hold on plot(yhat_n,'.-') hold off legend(["Measured" "Predicted"]) xlabel("Timestep (15-minutes)") ylabel("Electrical Load (kW)") title(["Forecast using FFNN";"Compressor 1 - Non-Working Day"]) subplot(2,2,4) stem(yhat_n - y_n) xlabel("Timestep (15-minutes)") ylabel("Error (kW)") title("RMSE = " + nonworkingDayRMSE) writetable(T,'results/ac1_ffnn_metrics.csv','WriteRowNames',true) saveas(gcf, 'results/ac1_ffnn_result.png')
🌈3 Matlab代码实现
🎉4 参考文献
部分理论来源于网络，如有侵权请联系删除。
[1]Da-Chun Wu, Babak Bahrami Asl, Ali Razban, Jie Chen (2020) Air Compressor Load Forecasting using Artificial Neural Network
