172 KiB
172 KiB
In [2]:
from math import sqrt from numpy import concatenate from matplotlib import pyplot import pandas as pd import numpy as np from sklearn.preprocessing import MinMaxScaler from sklearn.preprocessing import LabelEncoder from sklearn.metrics import mean_squared_error from tensorflow.keras import Sequential from tensorflow.keras.layers import Dense from tensorflow.keras.layers import LSTM from tensorflow.keras.layers import Dropout from sklearn.model_selection import train_test_split import matplotlib.pyplot as plt
这段代码是一个函数 time_series_to_supervised，它用于将时间序列数据转换为监督学习问题的数据集。下面是该函数的各个部分的含义：
data: 输入的时间序列数据，可以是列表或2D NumPy数组。 n_in: 作为输入的滞后观察数，即用多少个时间步的观察值作为输入。默认值为96，表示使用前96个时间步的观察值作为输入。 n_out: 作为输出的观测数量，即预测多少个时间步的观察值。默认值为1，表示预测未来1个时间步的观察值。 dropnan: 布尔值，表示是否删除具有NaN值的行。默认为True，即删除具有NaN值的行。 函数首先检查输入数据的维度，并初始化一些变量。然后，它创建一个新的DataFrame对象 df 来存储输入数据，并保存原始的列名。接着，它创建了两个空列表 cols 和 names，用于存储新的特征列和列名。
接下来，函数开始构建特征列和对应的列名。首先，它将原始的观察序列添加到 cols 列表中，并将其列名添加到 names 列表中。然后，它依次将滞后的观察序列添加到 cols 列表中，并构建相应的列名，格式为 (原始列名)(t-滞后时间)。这样就创建了输入特征的部分。
接着，函数开始构建输出特征的部分。它依次将未来的观察序列添加到 cols 列表中，并构建相应的列名，格式为 (原始列名)(t+未来时间)。
最后，函数将所有的特征列拼接在一起，构成一个新的DataFrame对象 agg。如果 dropnan 参数为True，则删除具有NaN值的行。最后，函数返回处理后的数据集 agg。
In [3]:
def time_series_to_supervised(data, n_in=96, n_out=1,dropnan=True): """ :param data:作为列表或2D NumPy数组的观察序列。需要。 :param n_in:作为输入的滞后观察数（X）。值可以在[1..len（数据）]之间可选。默认为1。 :param n_out:作为输出的观测数量（y）。值可以在[0..len（数据）]之间。可选的。默认为1。 :param dropnan:Boolean是否删除具有NaN值的行。可选的。默认为True。 :return: """ n_vars = 1 if type(data) is list else data.shape[1] df = pd.DataFrame(data) origNames = df.columns cols, names = list(), list() cols.append(df.shift(0)) names += [('%s' % origNames[j]) for j in range(n_vars)] n_in = max(0, n_in) for i in range(n_in, 0, -1): time = '(t-%d)' % i cols.append(df.shift(i)) names += [('%s%s' % (origNames[j], time)) for j in range(n_vars)] n_out = max(n_out, 0) for i in range(1, n_out+1): time = '(t+%d)' % i cols.append(df.shift(-i)) names += [('%s%s' % (origNames[j], time)) for j in range(n_vars)] agg = pd.concat(cols, axis=1) agg.columns = names if dropnan: agg.dropna(inplace=True) return agg
In [4]:
# 加载数据 path1 = r"D:\project\小论文1-基于ICEEMDAN分解的时序高维变化的短期光伏功率预测模型\CEEMAN-PosConv1dbiLSTM-LSTM\模型代码流程\data6.csv"#数据所在路径 #我的数据是excel表，若是csv文件用pandas的read_csv()函数替换即可。 datas1 = pd.DataFrame(pd.read_csv(path1)) #我只取了data表里的第3、23、16、17、18、19、20、21、27列，如果取全部列的话这一行可以去掉 # data1 = datas1.iloc[:,np.r_[3,23,16:22,27]] data1=datas1.interpolate() values1 = data1.values print(data1.head()) print(data1.shape)
Temp Humidity GHI DHI Rainfall Power 0 19.779453 40.025826 3.232706 1.690531 0.0 0.0 1 19.714937 39.605961 3.194991 1.576346 0.0 0.0 2 19.549330 39.608631 3.070866 1.576157 0.0 0.0 3 19.405870 39.680702 3.038623 1.482489 0.0 0.0 4 19.387363 39.319881 2.656474 1.134153 0.0 0.0 (104256, 6)
In [6]:
# 使用MinMaxScaler进行归一化 scaler = MinMaxScaler(feature_range=(0, 1)) scaledData1 = scaler.fit_transform(data1) print(scaledData1.shape)
(104256, 6)
In [7]:
n_steps_in =96 #历史时间长度 n_steps_out=1#预测时间长度 processedData1 = time_series_to_supervised(scaledData1,n_steps_in,n_steps_out) print(processedData1.head())
0 1 2 3 4 5 0(t-96) \
96 0.555631 0.349673 0.190042 0.040558 0.0 0.236302 0.490360
97 0.564819 0.315350 0.211335 0.044613 0.0 0.258204 0.489088
98 0.576854 0.288321 0.229657 0.047549 0.0 0.279860 0.485824
99 0.581973 0.268243 0.247775 0.053347 0.0 0.301336 0.482997
100 0.586026 0.264586 0.266058 0.057351 0.0 0.322851 0.482632
1(t-96) 2(t-96) 3(t-96) ... 2(t-1) 3(t-1) 4(t-1) 5(t-1) \
96 0.369105 0.002088 0.002013 ... 0.166009 0.036794 0.0 0.214129
97 0.364859 0.002061 0.001839 ... 0.190042 0.040558 0.0 0.236302
98 0.364886 0.001973 0.001839 ... 0.211335 0.044613 0.0 0.258204
99 0.365615 0.001950 0.001697 ... 0.229657 0.047549 0.0 0.279860
100 0.361965 0.001679 0.001167 ... 0.247775 0.053347 0.0 0.301336
0(t+1) 1(t+1) 2(t+1) 3(t+1) 4(t+1) 5(t+1)
96 0.564819 0.315350 0.211335 0.044613 0.0 0.258204
97 0.576854 0.288321 0.229657 0.047549 0.0 0.279860
98 0.581973 0.268243 0.247775 0.053347 0.0 0.301336
99 0.586026 0.264586 0.266058 0.057351 0.0 0.322851
100 0.590772 0.258790 0.282900 0.060958 0.0 0.343360
[5 rows x 588 columns]
In [8]:
data_x = processedData1.loc[:,'0(t-96)':'5(t-1)']#去除power剩下的做标签列 data_y = processedData1.loc[:,'5']
冒号
In [9]:
data_x.shape
Out[9]:
(104159, 576)
In [10]:
data_y
Out[10]:
96 0.236302
97 0.258204
98 0.279860
99 0.301336
100 0.322851
...
104250 0.000000
104251 0.000000
104252 0.000000
104253 0.000000
104254 0.000000
Name: 5, Length: 104159, dtype: float64
In [11]:
data_y.shape
Out[11]:
(104159,)
In [33]:
# 计算训练集、验证集和测试集的大小 train_size = int(len(data_x) * 0.90) test_size = int(len(data_x) * 0.015) val_size = len(data_x) - train_size - test_size # 计算训练集、验证集和测试集的索引范围 train_indices = range(train_size) val_indices = range(train_size, train_size + val_size) test_indices = range(train_size + val_size, len(data_x)) # 根据索引范围划分数据集 train_X1 = data_x.iloc[train_indices].values.reshape((-1, n_steps_in, scaledData1.shape[1])) val_X1 = data_x.iloc[val_indices].values.reshape((-1, n_steps_in, scaledData1.shape[1])) test_X1 = data_x.iloc[test_indices].values.reshape((-1, n_steps_in, scaledData1.shape[1])) train_y = data_y.iloc[train_indices].values val_y = data_y.iloc[val_indices].values test_y = data_y.iloc[test_indices].values # reshape input to be 3D [samples, timesteps, features] train_X = train_X1.reshape((train_X1.shape[0], n_steps_in, scaledData1.shape[1])) val_X = val_X1.reshape((val_X1.shape[0], n_steps_in, scaledData1.shape[1])) test_X = test_X1.reshape((test_X1.shape[0], n_steps_in, scaledData1.shape[1])) print(train_X.shape, train_y.shape, val_X.shape, val_y.shape, test_X.shape, test_y.shape)
(93743, 96, 6) (93743,) (8854, 96, 6) (8854,) (1562, 96, 6) (1562,)
In [34]:
train_X1.shape
Out[34]:
(93743, 96, 6)
In [77]:
import tensorflow as tf from tensorflow.keras.layers import Input, Conv1D, Bidirectional, GlobalAveragePooling1D, Dense, GRU, MaxPooling1D from tensorflow.keras.models import Model from tensorflow.keras.initializers import RandomUniform class AttentionWithImproveRelativePositionEncoding(tf.keras.layers.Layer): def __init__(self, d_model, num_heads, max_len=5000): super(AttentionWithImproveRelativePositionEncoding, self).__init__() self.num_heads = num_heads self.d_model = d_model self.max_len = max_len self.wq = tf.keras.layers.Dense(d_model) self.wk = tf.keras.layers.Dense(d_model) self.wv = tf.keras.layers.Dense(d_model) self.dense = tf.keras.layers.Dense(d_model) self.position_encoding = ImproveRelativePositionEncoding(d_model) def call(self, v, k, q, mask): batch_size = tf.shape(q)[0] q = self.wq(q) k = self.wk(k) v = self.wv(v) # 添加位置编码 k += self.position_encoding (k) q += self.position_encoding (q) q = self.split_heads(q, batch_size) k = self.split_heads(k, batch_size) v = self.split_heads(v, batch_size) scaled_attention, attention_weights = self.scaled_dot_product_attention(q, k, v, mask) scaled_attention = tf.transpose(scaled_attention, perm=[0, 2, 1, 3]) concat_attention = tf.reshape(scaled_attention, (batch_size, -1, self.d_model)) output = self.dense(concat_attention) return output, attention_weights def split_heads(self, x, batch_size): x = tf.reshape(x, (batch_size, -1, self.num_heads, self.d_model // self.num_heads)) return tf.transpose(x, perm=[0, 2, 1, 3]) def scaled_dot_product_attention(self, q, k, v, mask): matmul_qk = tf.matmul(q, k, transpose_b=True) dk = tf.cast(tf.shape(k)[-1], tf.float32) scaled_attention_logits = matmul_qk / tf.math.sqrt(dk) if mask is not None: scaled_attention_logits += (mask * -1e9) attention_weights = tf.nn.softmax(scaled_attention_logits, axis=-1) output = tf.matmul(attention_weights, v) return output, attention_weights class ImproveRelativePositionEncoding(tf.keras.layers.Layer): def __init__(self, d_model, max_len=5000): super(ImproveRelativePositionEncoding, self).__init__() self.max_len = max_len self.d_model = d_model # 引入可变化的参数u和v进行线性变化 self.u = self.add_weight(shape=(self.d_model,), initializer=RandomUniform(), trainable=True) self.v = self.add_weight(shape=(self.d_model,), initializer=RandomUniform(), trainable=True) def call(self, inputs): seq_length = inputs.shape[1] pos_encoding = self.relative_positional_encoding(seq_length, self.d_model) # 调整原始的相对位置编码公式，将u和v参数融入其中 pe_with_params = pos_encoding * self.u+ pos_encoding * self.v return inputs + pe_with_params def relative_positional_encoding(self, position, d_model): pos = tf.range(position, dtype=tf.float32) i = tf.range(d_model, dtype=tf.float32) angles = 1 / tf.pow(10000.0, (2 * (i // 2)) / tf.cast(d_model, tf.float32)) angle_rads = tf.einsum('i,j->ij', pos, angles) #保留了sinous机制 # Apply sin to even indices; 2i angle_rads_sin = tf.sin(angle_rads[:, 0::2]) # Apply cos to odd indices; 2i+1 angle_rads_cos = tf.cos(angle_rads[:, 1::2]) pos_encoding = tf.stack([angle_rads_sin, angle_rads_cos], axis=2) pos_encoding = tf.reshape(pos_encoding, [1, position, d_model]) return pos_encoding def PosConv1biGRUWithSelfAttention(input_shape, gru_units, num_heads): inputs = Input(shape=input_shape) # CNN layer cnn_layer = Conv1D(filters=64, kernel_size=2, activation='relu')(inputs) cnn_layer = MaxPooling1D(pool_size=1)(cnn_layer) gru_output = Bidirectional(GRU(gru_units, return_sequences=True))(cnn_layer) # Apply Self-Attention self_attention =AttentionWithImproveRelativePositionEncoding(d_model=gru_units*2, num_heads=num_heads) gru_output, _ = self_attention(gru_output, gru_output, gru_output, mask=None) pool1 = GlobalAveragePooling1D()(gru_output) output = Dense(1)(pool1) return Model(inputs=inputs, outputs=output) input_shape = (96, 6) gru_units = 64 num_heads = 8 # Create model model = PosConv1biGRUWithSelfAttention(input_shape, gru_units, num_heads) model.compile(optimizer='adam', loss='mse') model.summary()
Model: "functional_4"
┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃ ┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩ │ input_layer_4 │ (None, 96, 6) │ 0 │ - │ │ (InputLayer) │ │ │ │ ├─────────────────────┼───────────────────┼────────────┼───────────────────┤ │ conv1d_4 (Conv1D) │ (None, 95, 64) │ 832 │ input_layer_4[0]… │ ├─────────────────────┼───────────────────┼────────────┼───────────────────┤ │ max_pooling1d_4 │ (None, 95, 64) │ 0 │ conv1d_4[0][0] │ │ (MaxPooling1D) │ │ │ │ ├─────────────────────┼───────────────────┼────────────┼───────────────────┤ │ bidirectional_4 │ (None, 95, 128) │ 49,920 │ max_pooling1d_4[… │ │ (Bidirectional) │ │ │ │ ├─────────────────────┼───────────────────┼────────────┼───────────────────┤ │ attention_with_imp… │ [(None, None, │ 66,304 │ bidirectional_4[… │ │ (AttentionWithImpr… │ 128), (None, 8, │ │ bidirectional_4[… │ │ │ None, None)] │ │ bidirectional_4[… │ ├─────────────────────┼───────────────────┼────────────┼───────────────────┤ │ global_average_poo… │ (None, 128) │ 0 │ attention_with_i… │ │ (GlobalAveragePool… │ │ │ │ ├─────────────────────┼───────────────────┼────────────┼───────────────────┤ │ dense_24 (Dense) │ (None, 1) │ 129 │ global_average_p… │ └─────────────────────┴───────────────────┴────────────┴───────────────────┘
Total params: 117,185 (457.75 KB)
Trainable params: 117,185 (457.75 KB)
Non-trainable params: 0 (0.00 B)
In [61]:
# Compile and train the model model.compile(optimizer='adam', loss='mean_squared_error') from keras.callbacks import EarlyStopping, ModelCheckpoint # 定义早停机制 early_stopping = EarlyStopping(monitor='val_loss', min_delta=0, patience=10, verbose=0, mode='min') # 拟合模型，并添加早停机制和模型检查点 history = model.fit(train_X, train_y, epochs=100, batch_size=64, validation_data=(val_X, val_y), callbacks=[early_stopping])
Epoch 1/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 106s 71ms/step - loss: 0.0198 - val_loss: 0.0016 Epoch 2/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 107s 73ms/step - loss: 0.0016 - val_loss: 0.0015 Epoch 3/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 108s 74ms/step - loss: 0.0015 - val_loss: 0.0015 Epoch 4/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 107s 73ms/step - loss: 0.0015 - val_loss: 0.0014 Epoch 5/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 106s 73ms/step - loss: 0.0014 - val_loss: 0.0016 Epoch 6/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 105s 71ms/step - loss: 0.0014 - val_loss: 0.0015 Epoch 7/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 104s 71ms/step - loss: 0.0014 - val_loss: 0.0014 Epoch 8/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 107s 73ms/step - loss: 0.0013 - val_loss: 0.0014 Epoch 9/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 107s 73ms/step - loss: 0.0013 - val_loss: 0.0014 Epoch 10/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 106s 72ms/step - loss: 0.0013 - val_loss: 0.0015 Epoch 11/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 105s 71ms/step - loss: 0.0013 - val_loss: 0.0014 Epoch 12/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 105s 72ms/step - loss: 0.0013 - val_loss: 0.0015 Epoch 13/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 107s 73ms/step - loss: 0.0013 - val_loss: 0.0014 Epoch 14/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 108s 74ms/step - loss: 0.0012 - val_loss: 0.0014 Epoch 15/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 107s 73ms/step - loss: 0.0013 - val_loss: 0.0014 Epoch 16/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 104s 71ms/step - loss: 0.0013 - val_loss: 0.0013 Epoch 17/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 105s 72ms/step - loss: 0.0013 - val_loss: 0.0013 Epoch 18/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 105s 72ms/step - loss: 0.0012 - val_loss: 0.0014 Epoch 19/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 102s 69ms/step - loss: 0.0012 - val_loss: 0.0013 Epoch 20/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 102s 70ms/step - loss: 0.0012 - val_loss: 0.0014 Epoch 21/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 103s 70ms/step - loss: 0.0012 - val_loss: 0.0014 Epoch 22/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 102s 70ms/step - loss: 0.0011 - val_loss: 0.0014 Epoch 23/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 102s 69ms/step - loss: 0.0012 - val_loss: 0.0018 Epoch 24/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 101s 69ms/step - loss: 0.0012 - val_loss: 0.0014 Epoch 25/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 102s 70ms/step - loss: 0.0011 - val_loss: 0.0014 Epoch 26/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 102s 70ms/step - loss: 0.0012 - val_loss: 0.0015 Epoch 27/100 1465/1465 ━━━━━━━━━━━━━━━━━━━━ 97s 66ms/step - loss: 0.0012 - val_loss: 0.0015
In [62]:
# 预测 lstm_pred = model.predict(test_X) # 将预测结果的形状修改为与原始数据相同的形状
49/49 ━━━━━━━━━━━━━━━━━━━━ 1s 16ms/step
In [63]:
test_y_pre=test_y
In [64]:
plt.plot(history.history['loss'], label='train') plt.plot(history.history['val_loss'], label='test') plt.legend() plt.show()
In [65]:
lstm_pred.shape
Out[65]:
(1562, 1)
In [66]:
test_y_pre.shape
Out[66]:
(1562,)
In [67]:
test_y_pre1=test_y_pre.reshape(1562,1)
In [68]:
test_y_pre1
Out[68]:
array([[0.90540195],
[0.90466702],
[0.89696645],
...,
[0. ],
[0. ],
[0. ]])
In [69]:
results1 = np.broadcast_to(lstm_pred, (1562, 6))
In [70]:
test_y2 = np.broadcast_to(test_y_pre1, (1562, 6))
In [71]:
# 反归一化 inv_forecast_y = scaler.inverse_transform(results1) inv_test_y = scaler.inverse_transform(test_y2)
In [72]:
inv_test_y
Out[72]:
array([[ 4.08374339e+01, 9.30529669e+01, 1.27546074e+03,
5.95908965e+02, 2.15485743e+01, 4.67959929e+00],
[ 4.08001461e+01, 9.29803001e+01, 1.27442567e+03,
5.95425556e+02, 2.15310831e+01, 4.67580080e+00],
[ 4.04094426e+01, 9.22188951e+01, 1.26358018e+03,
5.90360385e+02, 2.13478095e+01, 4.63600016e+00],
...,
[-5.09990072e+00, 3.53003502e+00, 2.91584611e-01,
3.66558254e-01, 0.00000000e+00, 0.00000000e+00],
[-5.09990072e+00, 3.53003502e+00, 2.91584611e-01,
3.66558254e-01, 0.00000000e+00, 0.00000000e+00],
[-5.09990072e+00, 3.53003502e+00, 2.91584611e-01,
3.66558254e-01, 0.00000000e+00, 0.00000000e+00]])
In [73]:
# 计算均方根误差 rmse = sqrt(mean_squared_error(inv_test_y[:,5], inv_forecast_y[:,5])) print('Test RMSE: %.3f' % rmse) #画图 plt.figure(figsize=(16,8)) plt.plot(inv_test_y[:,5], label='true') plt.plot(inv_forecast_y[:,5], label='pre') plt.legend() plt.show()
Test RMSE: 0.063
In [80]:
from sklearn.metrics import mean_squared_error, mean_absolute_error # 评价指标 # 使用sklearn调用衡量线性回归的MSE 、 RMSE、 MAE、r2 from math import sqrt from sklearn.metrics import mean_absolute_error from sklearn.metrics import mean_squared_error from sklearn.metrics import r2_score print('mean_squared_error:', mean_squared_error(lstm_pred, test_y_pre)) # mse) print("mean_absolute_error:", mean_absolute_error(lstm_pred, test_y_pre)) # mae print("rmse:", sqrt(mean_squared_error(lstm_pred,test_y_pre))) #r2对比区域 print("r2 score:", r2_score(inv_test_y[:], inv_forecast_y[:]))#预测50天数据
mean_squared_error: 0.00014629570256978046 mean_absolute_error: 0.008445659571024366 rmse: 0.01209527604355438 r2 score: 0.9988370101682903
In [75]:
df1 = pd.DataFrame(inv_test_y[:,5], columns=['column_name'])
In [58]:
# 指定文件路径和文件名，保存DataFrame到CSV文件中 df1.to_csv('test.csv', index=False)
In [45]:
df2 = pd.DataFrame(inv_forecast_y[:], columns=['column_name'])
In [46]:
# 指定文件路径和文件名，保存DataFrame到CSV文件中 df2.to_csv('forecast.csv', index=False)
In [ ]: