Neural network based prediction of competition behaviour in energy markets

A system, comprising:

• a memory storing instructions;
• one or more communication interfaces;
• and
• one or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to:
• process in a neural network, an input request to predict a plurality of bid quantities of one or more energy generators for a given time slot and a given day of a particular week;
• obtain historical information across a pre-defined time period corresponding to the one or more energy generators, the historical information comprising a demand, a clearing price, number of bids placed by the one or more energy generators, and an actual quantity of generation cleared by a market operator for the one or more energy generators, wherein the historical information is arranged as (i) a target bid of the given time slot and the given day of the given week and (ii) information for prediction specific to the one or more energy generators for the given time slot of (a) a previous day, and (b) same day of a relative previous week with reference to the given day of the given week, wherein the clearing price refers to a per unit price that the market operator provides to the energy generator whose bids have been cleared;
• train the neural network using (i) a feedforward structure with softmax output layer of the neural network, (ii) the rearranged historical information, and (iii) a sum absolute error loss employed with the one or more energy generators, to obtain a learnt set of weights, wherein the feedforward structure is explored with and without hidden layers with a 10-output softmax function at an output layer ensures that outputs are all non-negative lying between 0 and 1, and constrained to add up to 1, wherein the usage of softmax output layer is used for a class of generators whose bid quantity is constant, wherein the sum of absolute error is utilized such that same loss is employed for all energy generators in which the information differs from one energy generator to another energy generator, wherein the sum of absolute error value is proven to be less than and equal to 2;
• receive for prediction of bid quantity of one or more energy generators, most recent data pertaining to the given time slot of the given day of the particular week, wherein the most recent data comprises demand, clearing price, number of bids placed by of the one or more energy generators, and cleared quantity of the one or more energy generators;
• apply the learnt set of weights on the most recent data to predict a set of fractional bids associated with the one or more energy generators for the given time slot of the given day of the particular week, wherein the fractional bids refer to bids distributed among a plurality of price bands;
• and
• predict a normalized bid vector, instead of predicting an actual bid quantity, such that the sum of all normalized or the set of fractional bids add up to 1 and learns posterior probabilities as in a multiclass logistic regression setting, further uses a general loss matrix Lij, instead of using a 0-1 loss matrix between class labels, where (i, j)th element denotes loss incurred in predicting class ‘j’ while true class is ‘i’, and a final prediction is obtained by minimizing expected loss under the general loss matrix.

One or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause predicting competition behavior in energy markets by:

• receiving in a neural network, via one or more hardware processors, an input request to predict a plurality of bid quantities of one or more energy generators for a given time slot and a given day of a particular week;
• obtaining, via the one or more hardware processors, historical information across a pre-defined time period corresponding to the one or more energy generators, the historical information comprising a demand, a clearing price, number of bids placed by the one or more energy generators, and an actual quantity of generation cleared by a market operator for the one or more energy generators, wherein the historical information is arranged as (i) a target bid of the given time slot and the given day of the given week and (ii) information for prediction specific to the one or more energy generators for the given time slot of (a) a previous day, and (b) same day of a relative previous week with reference to the given day of the given week, wherein the clearing price refers to a per unit price that the market operator provides to the energy generator whose bids have been cleared;
• training the neural network, via the one or more hardware processors, using (i) a feedforward structure with softmax output layer, (ii) the rearranged historical information, and (iii) a sum absolute error loss employed with the one or more energy generators, to obtain a learnt set of weights, wherein the feedforward structure is explored with and without hidden layers with a 10-output softmax function at an output layer ensures that outputs are all non-negative lying between 0 and 1, and constrained to add up to 1, wherein the usage of softmax output layer is used for a class of generators whose bid quantity is constant, wherein the sum of absolute error is utilized such that same loss is employed for all energy generators in which the information differs from one energy generator to another energy generator, wherein the sum of absolute error value is proven to be less than and equal to 2;
• receiving, via the one or more hardware processors, for prediction of bid quantity of one or more energy generators, most recent data pertaining to the given time slot of the given day of the particular week, wherein the most recent data comprises demand, clearing price, number of bids placed by of the one or more energy generators, and cleared quantity of the one or more energy generators;
• applying the learnt set of weights on the most recent data to predict a set of fractional bids associated with the one or more energy generators for the given time slot of the given day of the particular week, wherein the fractional bids refer to bids distributed among a plurality of price bands;
• and
• predicting a normalized bid vector, instead of predicting an actual bid quantity, such that the sum of all normalized or the set of fractional bids add up to 1 and learning posterior probabilities as in a multiclass logistic regression setting, further uses a general loss matrix Lij, instead of using a 0-1 loss matrix between class labels, where (i, j)th element denotes loss incurred in predicting class ‘j’ while true class is ‘i’, and a final prediction is obtained by minimizing expected loss under the general loss matrix.