The following database named EVA keeps a running tally of the effectiveness
of various secondary structure prediction schemes.
http://cubic.bioc.columbia.edu/eva/index.html
You can find links and references for a variety of advanced methods there.
All of the good methods achieve their power by first aligning homologous sequences so that the prediction is based on consensus information instead of on a single sequence. One of these methods should be used instead of any algorithm you might find in a package that works over just one sequence.
Of the best of methods available over the web are PsiPred, and SAM. PsiPred takes your input sequence sna derives the other members of the family that will be aligned with it by a PsiBlast search. SAM takes you input sequence and does an HMM search to create the aligned family followed by a neural net analysis to predict the secondary structure.
There is a version of PsiPred running at the bioinformatics center that can be run with greater flexibility than the web PsiPred. It can be run in a batch mode for making large numbers of predictions. It can be configured to accept an alignment of your choosing rather than insisting on a PsiBlast alignment. The local SAM implementation lacks the neural net secondary structure predictor, but can be used to make HMM alignments for processing by PsiPred. The local SAM implementation can also make a secondary structure consensus over a set of sequences for which secondary structure predictions have already been provided by some other means.
Rost B, and Eyrich VA. 2001. EVA: large-scale analysis of secondary
structure prediction. Proteins. Suppl 5:192-199, 2001.