前言
-
LSTM 航空乘客预测单步预测的两种情况。 简单运用LSTM 模型进行预测分析。
-
加入注意力机制的LSTM 对航空乘客预测采用了目前市面上比较流行的注意力机制，将两者进行结合预测。
-
多层 LSTM 对航空乘客预测 简单运用多层的LSTM 模型进行预测分析。
-
双向LSTM 对航空乘客预测双向LSTM网络对其进行预测。
-
MLP多层感知器 对航空乘客预测简化版 使用MLP 对航空乘客预测
-
CNN + LSTM 航空乘客预测采用的CNN + LSTM网络对其进行预测。
-
ConvLSTM 航空乘客预测采用ConvLSTM 航空乘客预测
-
LSTM的输入格式和输出个数说明 中对单步和多步的输入输出格式进行了解释
- LSTM 单变量多步预测航空乘客简单版
- LSTM 单变量多步预测航空乘客复杂版
-
LSTM 多变量单步预测空气质量（1—》1） 用LSTM 前一个数据点的多变量预测下一个时间点的空气质量
-
LSTM 多变量单步预测空气质量（3 —》1） 用LSTM 前三个数据点的多变量预测下一个时间点的空气质量
本文主要是采用麻雀算法SSA优化LSTM超参数
程序
麻雀搜索算法是2020提出的一种新的优化算法,在此不对具体原理进行分析,针对代码实操.
SSA
麻雀算法代码简介
class SSA():
def __init__(self, func, n_dim=None, pop_size=20, max_iter=50, lb=-512, ub=512, verbose=False):
self.func = func
self.n_dim = n_dim # dimension of particles, which is the number of variables of func
self.pop = pop_size # number of particles
P_percent = 0.2 # # 生产者的人口规模占总人口规模的20%
D_percent = 0.1 # 预警者的人口规模占总人口规模的10%
self.pNum = round(self.pop * P_percent) # 生产者的人口规模占总人口规模的20%
self.warn = round(self.pop * D_percent) # 预警者的人口规模占总人口规模的10%
self.max_iter = max_iter # max iter
self.verbose = verbose # print the result of each iter or not
self.lb, self.ub = np.array(lb) * np.ones(self.n_dim), np.array(ub) * np.ones(self.n_dim)
assert self.n_dim == len(self.lb) == len(self.ub), 'dim == len(lb) == len(ub) is not True'
assert np.all(self.ub > self.lb), 'upper-bound must be greater than lower-bound'
self.X = np.random.uniform(low=self.lb, high=self.ub, size=(self.pop, self.n_dim))
self.Y = [self.func(self.X[i]) for i in range(len(self.X))] # y = f(x) for all particles
self.pbest_x = self.X.copy() # personal best location of every particle in history
self.pbest_y = [np.inf for i in range(self.pop)] # best image of every particle in history
self.gbest_x = self.pbest_x.mean(axis=0).reshape(1, -1) # global best location for all particles
self.gbest_y = np.inf # global best y for all particles
self.gbest_y_hist = [] # gbest_y of every iteration
self.update_pbest()
self.update_gbest()
#
# record verbose values
self.record_mode = False
self.record_value = {'X': [], 'V': [], 'Y': []}
self.best_x, self.best_y = self.gbest_x, self.gbest_y # history reasons, will be deprecated
self.idx_max = 0
self.x_max = self.X[self.idx_max, :]
self.y_max = self.Y[self.idx_max]
def cal_y(self, start, end):
# calculate y for every x in X
for i in range(start, end):
self.Y[i] = self.func(self.X[i])
# return self.Y
def update_pbest(self):
'''
personal best
'''
for i in range(len(self.Y)):
if self.pbest_y[i] > self.Y[i]:
self.pbest_x[i] = self.X[i]
self.pbest_y[i] = self.Y[i]
def update_gbest(self):
idx_min = self.pbest_y.index(min(self.pbest_y))
if self.gbest_y > self.pbest_y[idx_min]:
self.gbest_x = self.X[idx_min, :].copy()
self.gbest_y = self.pbest_y[idx_min]
def find_worst(self):
self.idx_max = self.Y.index(max(self.Y))
self.x_max = self.X[self.idx_max, :]
self.y_max = self.Y[self.idx_max]
def update_finder(self):
r2 = np.random.rand(1) # 预警值
self.idx = sorted(enumerate(self.Y), key=lambda x: x[1])
self.idx = [self.idx[i][0] for i in range(len(self.idx))]
# 这一部位为发现者（探索者）的位置更新
if r2 < 0.8: # 预警值较小，说明没有捕食者出现
for i in range(self.pNum):
r1 = np.random.rand(1)
self.X[self.idx[i], :] = self.X[self.idx[i], :] * np.exp(-(i) / (r1 * self.max_iter)) # 对自变量做一个随机变换
self.X = np.clip(self.X, self.lb, self.ub) # 对超过边界的变量进行去除
# X[idx[i], :] = Bounds(X[idx[i], :], lb, ub) # 对超过边界的变量进行去除
# fit[sortIndex[0, i], 0] = func(X[sortIndex[0, i], :]) # 算新的适应度值
elif r2 >= 0.8: # 预警值较大，说明有捕食者出现威胁到了种群的安全，需要去其它地方觅食
for i in range(self.pNum):
Q = np.random.rand(1) # 也可以替换成 np.random.normal(loc=0, scale=1.0, size=1)
self.X[self.idx[i], :] = self.X[self.idx[i], :] + Q * np.ones(
(1, self.n_dim)) # Q是服从正态分布的随机数。L表示一个1×d的矩阵
self.X = np.clip(self.X, self.lb, self.ub) # 对超过边界的变量进行去除
# X[idx[i], :] = Bounds(X[sortIndex[0, i], :], lb, ub)
# fit[sortIndex[0, i], 0] = func(X[sortIndex[0, i], :])
self.cal_y(0, self.pNum)
def update_follower(self):
# 这一部位为加入者（追随者）的位置更新
for ii in range(self.pop - self.pNum):
i = ii + self.pNum
A = np.floor(np.random.rand(1, self.n_dim) * 2) * 2 - 1
best_idx = self.Y[0:self.pNum].index(min(self.Y[0:self.pNum]))
bestXX = self.X[best_idx, :]
if i > self.pop / 2:
Q = np.random.rand(1)
self.X[self.idx[i], :] = Q * np.exp((self.x_max - self.X[self.idx[i], :]) / np.square(i))
else:
self.X[self.idx[i], :] = bestXX + np.dot(np.abs(self.X[self.idx[i], :] - bestXX),
1 / (A.T * np.dot(A, A.T))) * np.ones((1, self.n_dim))
self.X = np.clip(self.X, self.lb, self.ub) # 对超过边界的变量进行去除
# X[self.idx[i],:] = Bounds(X[self.idx[i],lb,ub)
# fit[self.idx[i],0] = func(X[self.idx[i], :])
self.cal_y(self.pNum, self.pop)
def detect(self):
arrc = np.arange(self.pop)
c = np.random.permutation(arrc) # 随机排列序列
b = [self.idx[i] for i in c[0: self.warn]]
e = 10e-10
for j in range(len(b)):
if self.Y[b[j]] > self.gbest_y:
self.X[b[j], :] = self.gbest_y + np.random.rand(1, self.n_dim) * np.abs(self.X[b[j], :] - self.gbest_y)
else:
self.X[b[j], :] = self.X[b[j], :] + (2 * np.random.rand(1) - 1) * np.abs(
self.X[b[j], :] - self.x_max) / (self.func(self.X[b[j]]) - self.y_max + e)
# X[sortIndex[0, b[j]], :] = Bounds(X[sortIndex[0, b[j]], :], lb, ub)
# fit[sortIndex[0, b[j]], 0] = func(X[sortIndex[0, b[j]]])
self.X = np.clip(self.X, self.lb, self.ub) # 对超过边界的变量进行去除
self.Y[b[j]] = self.func(self.X[b[j]])
def run(self, max_iter=None):
self.max_iter = max_iter or self.max_iter
for iter_num in range(self.max_iter):
self.update_finder() # 更新发现者位置
self.find_worst() # 取出最大的适应度值和最差适应度的X
self.update_follower() # 更新跟随着位置
self.update_pbest()
self.update_gbest()
self.detect()
self.update_pbest()
self.update_gbest()
self.gbest_y_hist.append(self.gbest_y)
return self.best_x, self.best_y
LSTM
def build_model(neurons1, neurons2, dropout):
X_train, y_train, X_test, y_test = process_data()
# X_train, y_train = create_dataset(X_train, y_train, steps)
# X_test, y_test = create_dataset(X_test, y_test, steps)
nb_features = X_train.shape[2]
input1 = X_train.shape[1]
model1 = Sequential()
model1.add(LSTM(
input_shape=(input1, nb_features),
units=neurons1,
return_sequences=True))
model1.add(Dropout(dropout))
model1.add(LSTM(
units=neurons2,
return_sequences=False))
model1.add(Dropout(dropout))
model1.add(Dense(units=1))
model1.add(Activation("linear"))
model1.compile(loss='mse', optimizer='Adam', metrics='mae')
return model1, X_train, y_train, X_test, y_test
优化超参数
if __name__ == '__main__':
'''
神经网络第一层神经元个数
神经网络第二层神经元个数
dropout比率
batch_size
'''
neurons1 = 64
neurons2 = 64
dropout = 0.01
batch_size = 32
model, X_train, y_train, X_test, y_test = build_model(neurons1, neurons2, dropout)
history1 = model.fit(X_train, y_train, epochs=150, batch_size=batch_size, validation_split=0.2, verbose=1,
callbacks=[EarlyStopping(monitor='val_loss', patience=9, restore_best_weights=True)])
# 测试集预测
y_score = model.predict(X_test)
# 反归一化
y_score = scaler.inverse_transform(y_score.reshape(-1, 1))
y_test = scaler.inverse_transform(y_test.reshape(-1, 1))
print("==========evaluation==============\n")
from sklearn.metrics import mean_squared_error
from sklearn.metrics import mean_absolute_error #平方绝对误差
import math
MAE = mean_absolute_error(y_test, y_score)
print('MAE: %.4f ' % MAE)
RMSE = math.sqrt(mean_squared_error(y_test, y_score))
print('RMSE: %.4f ' % (RMSE))
总结
- SSA在一定范围内可以优化LSTM 的超参数,对算力要求有点大
- SSA优化算法有一定的局限性,如何利用其优势至关重要
- LSTM的超参数可以部分优化,能够节约时间和节省算力资源
备注:
需要源代码和数据集,或者想要沟通交流,请私聊,谢谢.
