Accu­rate wind pow­er fore­cast­ing is essen­tial to reduce the neg­a­tive impact of wind power

on the oper­a­tion of the grid and the oper­a­tion cost of the pow­er sys­tem. Day-ahead wind power

fore­cast­ing plays an impor­tant role in the day-ahead elec­tric­i­ty spot trad­ing mar­ket. How­ev­er, the

insta­bil­i­ty of the wind pow­er series makes the fore­cast dif­fi­cult. To improve fore­cast accu­ra­cy, a

hybrid opti­miza­tion algo­rithm is estab­lished in this study, which com­bines vari­a­tion­al mode decomposition

(VMD), max­i­mum rel­e­vance & min­i­mum redun­dan­cy algo­rithm (mRMR), long shortterm

mem­o­ry neur­al net­work (LSTM), and fire­fly algo­rithm (FA) togeth­er. First­ly, the orig­i­nal historical

wind pow­er sequence is decom­posed into sev­er­al char­ac­ter­is­tic mod­el func­tions with VMD.

Then, mRMR is applied to obtain the best fea­ture set by ana­lyz­ing the cor­re­la­tion between each

com­po­nent. Final­ly, the FA is used to opti­mize the var­i­ous para­me­ters LSTM. Adding the forecasting

results of all sub-sequences acquires the fore­cast­ing result. It turns out that the pro­posed hybrid

algo­rithm is supe­ri­or to the oth­er six com­par­i­son algo­rithms. At the same time, an addi­tion­al case

is pro­vid­ed to fur­ther ver­i­fy the adapt­abil­i­ty and sta­bil­i­ty of the pro­posed hybrid model.