How to train multiple time series using one LSTM scriptt?

For example, I have electricity usage data from different cities as:

year :1 2 3 4 5 6 [years]

city 1 : 1 2 3 4 5 6 [GWh]

city 2 : 1 2 3 4 5 6 [GWh]

city 3 : 1 2 3 4 5 6 [GWh]

I want to predict for year 7. I can do thhat with narxnet or narnet by using a 3x6 matrix, but these two are inconsistant. LSTM seems to be mroe accurate.

I can predict for one city using the following script. P1 is electricity data for city 1.

%% clear variables

% clc; clear; close all

%% load data

data=p1;

%% partition data

trec=length(data);

trrec=0.9*trec;

NTST=round(trrec);

datatrain=data(1:NTST+1);

datatest=data(NTST+1:end);

%% standardize data

mu=mean(datatrain);

sig=std(datatrain);

datatrainstd=(datatrain-mu)/sig;

%% display standardized data

% figure,

% hold on; grid on; box on, grid minor

% set(gca,'FontSize',40)

% set(gca, 'FontName', 'Times New Roman')

% plot(datatrainstd)

% xlabel('Months')

% ylabel('Cases')

%% train network separate predictors and response

xtrain=datatrainstd(1:end-1);

ytrain=datatrainstd(2:end);

%% define lstm network architecture

nof=1;

nor=1;

nhu=10;

layers=[

sequenceInputLayer(nof,'Name','ip')

lstmLayer(nhu,"Name",'lstm')

fullyConnectedLayer(nor,'Name','fc')

regressionLayer('Name','rg')];

%% plot layers

% lgraph=layerGraph(layers);

% plot(lgraph)

%% training options

options=trainingOptions('adam',...

'MaxEpochs',200,...

'GradientThreshold',1,...

'InitialLearnRate',0.1,...

'LearnRateSchedule','piecewise',...

'LearnRateDropperiod',100,...

'LearnRateDropFactor',0.2,...

'Verbose',0,...

'plots','training-progres');

%% training lstm network

net=trainNetwork(xtrain,ytrain,layers,options);

%% validation

%% standardize validation data using same parameters of training phase

datateststd=(datatest-mu)/sig;

xtest=datateststd(1:end-1);

ytest=datatest(2:end);

net=predictAndUpdateState(net,xtrain);

[net, ypred] = predictAndUpdateState(net, ytrain(end));

NTSTs=length(xtest);

for i=2:NTSTs

[net, ypred(:,i)] = predictAndUpdateState(net, ytrain(:,i-1));

end

%% unstandarize data

ypred=sig*ypred+mu;

rmse=sqrt(mean(ypred-ytest).^2);

ase=mean(mode(ypred-ytest));

ASME=mode(mean((ypred-ytest)));

%% plotting

it=1:length(ypred);

figure,

hold on

plot(datatrain(1:end-1))

startpnt=NTST;

endpt=NTST+NTSTs;

idx=startpnt:endpt;

plot(idx,[data(startpnt) ypred],'o--')

plot(idx,[data(startpnt) ytest],'+--')

hold off

Answer 1

[Rahul]

Hi @Fahad Butt,

In order to train LSTM model over multiple time series data using a single script, consider representating all time series data in a single matrix like this:

data = [p1; p2; p3];
% Considering 'p1', 'p2', 'p3' contain the electricity data different cities
numCities = size(data, 1);
trec = size(data, 2);

Now, apply the same logic of training the LSTM as mentioned in the question. The difference would be to apply the logic iteratively using for loops. Here is an example:

nets = cell(1, numCities);
for i = 1:numCities
    nets{i} = trainNetwork(squeeze(xtrain(i, :)), squeeze(ytrain(i, :)), layers, options);
end

If the time series data from different cities share similar patterns that can be standardized collectively, then concatenating the data and training can be an effective option.

concatenatedData = data(:)'; % Convert matrix to a single row vector

Note: The concatenated sequence can become very long, which may require adjustments in model architecture or training options to handle effectively.

The following MathWorks documentations can be referred:

'LSTM Neural Networks': https://www.mathworks.com/help/deeplearning/ug/long-short-term-memory-networks.html

Thanks.